GitLab's Sijbrandij Uses AI to Build Personalized Cancer Treatment Plan
OpenAI
Combining artificial intelligence models with open genomic data allows terminal patients to bypass traditional clinical trial bottlenecks and design targeted multi-drug therapies.
## Background
Published by OpenAI on YouTube on April 9, 2026, this OpenAI Forum presentation features GitLab co-founder Sid Sijbrandij and geneticist Jacob Stern discussing how they approached Sid's rare, aggressive cancer diagnosis like an engineering problem. The page description outlines how the pair leveraged ChatGPT, custom diagnostics, and real-time scientific literature to build a hyper-fast feedback loop between patients, clinicians, and cutting-edge biotech research.
## TL;DR
Faced with a terminal recurrence of osteosarcoma and no remaining standard treatment options, Sid Sijbrandij adopted a high-intensity 'founder mode' strategy, publishing 25 terabytes of his diagnostic data publicly and funding personalized therapeutic pathways. Collaborating with geneticist Jacob Stern, the team utilized frontier AI models and autonomous multi-agent bioinformatics pipelines to analyze hundreds of thousands of single-cell datasets, identify rare overexpressed targets, and bypass off-target organ toxicities using engineered AND-logic gated CAR-Ts and bespoke mRNA vaccines. By combining first-principles reasoning, targeted radiopharmaceuticals, and global trial sourcing, Sid successfully expanded his viable therapeutic options from zero to over thirty backup treatments while achieving a disease-free status.
## Atmosphere & style
The session balances the tense gravity of a life-or-death medical crisis with the pragmatic, highly disciplined focus of Silicon Valley engineering. Sid and Jacob speak with a calm, analytical candor that renders complex biotech jargon accessible, creating a narrative arc that moves from shocking diagnostic scares to empowering technological breakthroughs. The tone is intellectually rigorous yet deeply human, conveying a sense of quiet urgency, optimism, and radical patient agency.
## Key takeaways
- [00:09](https://www.youtube.com/watch?v=OAlHiQLsYQM&t=9s) GitLab co-founder Sid Sijbrandij and geneticist Jacob Stern adopted a high-intensity, data-driven approach using ChatGPT and advanced diagnostics to manage Sid's rare spine tumor.
- [03:05](https://www.youtube.com/watch?v=OAlHiQLsYQM&t=185s) After experiencing recurrence with limited standard-of-care options, Sid uploaded 25 terabytes of his own diagnostic and genomic data to osteosarc.com for open scientific research.
- [08:21](https://www.youtube.com/watch?v=OAlHiQLsYQM&t=501s) Using single-cell sequencing to identify FAP expression, Sid underwent targeted radiopharmaceutical therapy in Germany, achieving 60% tumor necrosis and allowing successful surgical resection.
- [10:50](https://www.youtube.com/watch?v=OAlHiQLsYQM&t=650s) Jacob Stern built multi-agent AI frameworks that execute Python code and bioinformatic workflows, enabling rapid analysis of 600,000 single-cell data points across longitudinal blood draws.
- [16:54](https://www.youtube.com/watch?v=OAlHiQLsYQM&t=1014s) To prevent fatal liver toxicity discovered via a B7-H3 PET scan in Beijing, the team developed an AND-logic gated CAR-T therapy requiring dual expression of B7-H3 and FAP.
- [23:38](https://www.youtube.com/watch?v=OAlHiQLsYQM&t=1418s) AI acts as a force multiplier and 'Iron Man suit' for non-specialists, democratizing complex technical knowledge and accelerating bespoke mRNA vaccine production to six months.
- [26:54](https://www.youtube.com/watch?v=OAlHiQLsYQM&t=1614s) Sid argues that oncologists prioritize minimizing legal liability, whereas terminal patients prioritize maximizing survival through aggressive, AI-designed multi-drug combinations.
- [26:54](https://www.youtube.com/watch?v=OAlHiQLsYQM&t=1614s) Basic high-value cancer diagnostics like RNA sequencing (~$50) and whole genome sequencing (~$500) are surprisingly affordable, lowering financial barriers to personalized data collection.
- [30:01](https://www.youtube.com/watch?v=OAlHiQLsYQM&t=1801s) Through global sourcing, advanced diagnostics, and AI analysis, Sid expanded his personal therapeutic ladder from zero standard options to over 30 candidate backup treatments.
- [30:01](https://www.youtube.com/watch?v=OAlHiQLsYQM&t=1801s) Sid and Jacob launched biotech startups Valis and Arden to scale personalized oncology and automated immunoprofiling workflows using AI as a deflationary engine.
- [36:00](https://www.youtube.com/watch?v=OAlHiQLsYQM&t=2160s) A terrifying misdiagnosis of inoperable lung metastasis was later identified as COVID-19 tissue remnants, highlighting the need for AI-driven differential diagnostic tools.
- [42:40](https://www.youtube.com/watch?v=OAlHiQLsYQM&t=2560s) High drug development costs ($1B+) and administrative IRB delays bottleneck progress, prompting calls for streamlined trial frameworks like Australia's notification system and 'IRB freedom'.
## Detailed walkthrough
### [00:09](https://www.youtube.com/watch?v=OAlHiQLsYQM&t=9s) Introduction to the OpenAI Forum and Sid Sijbrandij's Story
Chris Nicholson opens the OpenAI Forum, introducing it as a global community bringing leaders together to share real-world AI applications, key learnings, and ethical impact. He introduces two featured speakers: Sid Sijbrandij, the co-founder and executive chair of GitLab, and Jacob Stern, a geneticist collaborating directly with Sid. Chris notes his decade-long personal friendship with Sid before turning the floor over to Scott McKinney, an OpenAI researcher.
Scott shares his own battle with osteosarcoma—a rare, aggressive bone cancer affecting fewer than 1,000 Americans annually. He describes standard-of-care treatments as sparse, blunt, and gruesome, emphasizing that survival upon recurrence is typically measured in months. Scott highlights how Sid and Jacob adopted a high-intensity "founder mode" approach to craft personalized therapies after exhausting conventional clinical options. Connecting with Sid and Jacob during his own diagnosis, Scott notes that their shared playbook drastically altered how he navigated his own condition.
- Osteosarcoma strikes fewer than 1,000 Americans per year, with extremely limited post-recurrence options.
- OpenAI is actively expanding AI in healthcare through public model benchmarks, clinical co-pilots in primary care, and ChatGPT Health.
[03:05](https://www.youtube.com/watch?v=OAlHiQLsYQM&t=185s) Diagnosis, Recurrence, and the Shift to "Founder Mode"
Sid Sijbrandij details his initial diagnosis and medical background, starting with his 2015 move to the U.S. to build GitLab and taking it public six years later. In 2022, while doing a bench press, he experienced chest pain around his heart. Initial emergency room X-rays showed no cardiac issues, but hours later his general practitioner urgently called asking, "Do you know how to meditate?" The doctor warned him that extreme blood pressure could indicate a bursting aorta. Subsequent emergency imaging cleared his aorta but revealed a 6 cm tumor growing from his vertebrae.
Sid underwent immediate spine fusion surgery, hardware placement, intensive radiation, and heavy chemotherapy that required four blood transfusions to survive. Around the same time, the 2022 Nobel Prize in Chemistry was awarded for click chemistry—a modular technique allowing two chemical compounds to connect cleanly inside a human body without side effects. Sid had previously met founder Jose during Y Combinator and had become the primary investor in Jose's click-chemistry biotech startup. Recognizing Sid's critical status, Jose helped set up a custom trial under an expanded access pathway. Sid emphasizes the FDA's single-patient Investigational New Drug (IND) pathway as a vital option for rare conditions.
> "He's like, 'Well, you better start meditating right now because you have super high blood pressure and you might have an aneurysm. And the pain you're feeling might be your aorta starting to burst. So, chill out and also get to the emergency room right now.'"
Despite these aggressive efforts, the cancer recurred two years later as a local progression. With no remaining standard-of-care options or applicable clinical trials, Sid stepped down from his daily routine to focus fully on managing his cancer. He adopted a parallelized approach to diagnostics and treatment development, publishing 25 terabytes of his own clinical and genomic data publicly at osteosarc.com.
- Single-patient INDs have near-total approval rates by the FDA (99.7% to 100%), offering a viable path for customized treatments.
- Max-data approach: Sid uploaded 25 terabytes of personal tumor diagnostic data to osteosarc.com to assist researchers worldwide.
[08:21](https://www.youtube.com/watch?v=OAlHiQLsYQM&t=501s) Targeted Radiotherapy, Tumor Microenvironment, and MDM2
Through single-cell sequencing, Sid's team identified high expressions of Fibroblast Activation Protein (FAP) on his cancer cells. Sid traveled to Germany for an experimental targeted radiopharmaceutical therapy combining FAP-targeting molecules with radioactive isotopes. The treatment achieved a 60% tumor necrosis rate and 20% shrinkage, detaching the tumor from his dura and allowing surgeons to remove it cleanly.
Sid also found extreme overexpression of the MDM2 protein in his tumor, placing him in the top 0.3% of index cases. While pharmaceutical companies had created MDM2 inhibitor drugs, development was halted due to narrow commercial potential. Sid stepped in to fund the maintenance of frozen trial samples and actively seeks ways to bring these therapies to market for the minority of patients who could benefit.
- Single-cell sequencing and TCR (T-cell receptor) tracking confirmed strong T-cell infiltration following target modulation.
- Sid is personally paying to preserve abandoned MDM2 inhibitor drug inventories to keep clinical access open.
[10:50](https://www.youtube.com/watch?v=OAlHiQLsYQM&t=650s) AI-Driven Data Analysis and Multi-Agent Workflows
Jacob Stern describes joining the effort after meeting Sid through their mutual friend Jose. Drawing on his six years at 10x Genomics, Jacob began integrating multi-modal diagnostics, including single-cell sequencing, bulk RNA/DNA profiling, targeted radiodiagnostics, tissue staining, and patient-derived organoid drug testing. To bridge his knowledge gaps across these domains, Jacob relies heavily on AI models to analyze data and interact with specialist clinicians.
```
Natural Language Prompt
│
▼
Spin up LLM Agents (Literature Search & Bioinformatic Hypotheses)
│
▼
Execute Python Analysis against 600,000 Single-Cell Data Points
│
▼
Synthesize Findings into Interactive Plots & Code Output
```
Jacob shares how early tests with ChatGPT (using GPT-4 on the Pro plan) successfully identified key expression markers like B7-H3 directly from CSV files of bulk RNA sequencing counts. He then details a multi-agent framework they built to run natural language bioinformatics workflows. In one case, after receiving a warning about possible Clonal Hematopoiesis of Indeterminate Potential (CHIP)—a blood condition linked to previous chemotherapy exposure—Jacob queried the agent network. The system ran literature searches, selected surface markers, analyzed 600,000 single-cell data points across multiple blood draws, and generated interactive plots alongside Python code over a 30-minute reasoning run, ultimately clearing the risk.
- AI agents processed 600k individual blood cells across multiple longitudinal time points in 30 minutes.
- The pipeline automates literature review, Python code generation, bioinformatic tool execution, and visualization.
[15:38](https://www.youtube.com/watch?v=OAlHiQLsYQM&t=938s) De Novo Drug Development: mRNA Vaccines and TCR-T Therapies
Jacob introduces the custom therapies being produced specifically for Sid, describing the process of creating bespoke medicines outside traditional pharma pipelines as "radicalizing." He outlines two core modalities: custom mRNA cancer vaccines and T-cell receptor (TCR) T-cell therapies.
The personalized mRNA vaccine functions similarly to viral vaccines: instead of expressing viral antigens, it encodes specific neoantigen mutations unique to Sid's tumor cells, priming his immune system to recognize and target those foreign markers. Jacob then introduces TCR-T cell therapies as an engineered cell approach designed to equip T cells to hunt down intracellular tumor targets.
[16:54](https://www.youtube.com/watch?v=OAlHiQLsYQM&t=1014s) Next-Gen Immunotherapies: T-Cell Receptors (TCR-Ts) and Dual-Gated CAR-Ts
Jacob Stern elaborates on the downstream immunotherapies they have prepared as backup options. While vaccines train the immune system broadly, T-cell receptor (TCR) therapy is more direct. T-cells act as the "business end" of the immune system; they use specialized receptors to identify specific targets and literally punch holes in tumor cells by firing lethal proteins at them. Through single-cell sequencing and work with specialized partners, the team identified specific TCRs tailored to Sid’s tumor, which can be engineered into healthy T-cells and mass-produced for infusion if necessary.
Beyond TCR-Ts, the team developed a customized Chimeric Antigen Receptor T-cell (CAR-T) therapy—described as their ultimate "nuclear option" on the shelf. They initially picked B7-H3 as the target protein expressed across most of Sid's cancer cells. However, when Sid flew to Beijing to undergo a specialized PET scan to map B7-H3 expression, the doctors delivered a mix of good and bad news: no active cancer was visible, but his liver bound three and a half times more tracer than any of the control subjects. Giving a standard B7-H3 CAR-T could have completely destroyed his liver.
To navigate around this off-target liver toxicity, the team collaborated with researcher Cole Roybal, who pioneered "AND-logic gated" CAR-Ts. By requiring the engineered T-cell to detect *both* B7-H3 and Fibroblast Activation Protein (FAP)—a marker with extremely low liver expression—they ensured the "nuclear bomb" would only detonate inside tumor tissue while sparing healthy organs.
- **T-Cell Mechanics:** T-cells recognize ultra-specific targets via TCRs and release pore-forming proteins to destroy target cells.
- **B7-H3 Diagnostic Scan:** Travel to Beijing in October revealed high liver uptake (3.5x normal controls), identifying a fatal safety risk before treatment delivery.
- **AND-Logic Gate Solution:** Engineered by Cole Roybal; requires simultaneous expression of B7-H3 AND FAP to trigger cell killing.
- **Discovery of PENXIN3:** Genomic data revealed PENXIN3 was overexpressed in Sid's tumor by 10,000x compared to healthy tissue.
- **Hydrophobic Target Challenge:** Because PENXIN3 is hydrophobic, traditional water-based lab assays failed to identify it in historical literature. AI-driven data mining uncovered it, leading the team to launch a targeted binder discovery campaign.
> "T-cells are very killy... basically they shoot out proteins that rip holes in cells that are close to them."
> "The bad news is that my liver was showing up three and a half times as much as the 20 Chinese people they scanned before me. So now we were very worried. If I ever were to take this medicine, I might lose my liver."
---
[23:38](https://www.youtube.com/watch?v=OAlHiQLsYQM&t=1418s) AI as an "Iron Man Suit" for Navigating Arcane Medicine
Jacob opens up his personal ChatGPT chat history to illustrate how he acts as a non-specialist directing advanced biomedical projects. Two years prior, navigating expert discussions across various biotech operations and academic groups would have been impossible. ChatGPT’s "Deep Research" functions served as a force multiplier, democratizing specialized technical knowledge so Jacob could hold rigorous, peer-level conversations with world-class scientists without making basic errors.
When designing Sid’s personalized mRNA vaccine, the team had to decide whether to work with large Contract Development and Manufacturing Organizations (CDMOs)—the types of facilities that produced commercial COVID-19 vaccines—or an academic lab. Using AI, they reasoned through first-principles quality control (QC) metrics to ensure the custom formulation would be clean, effective, and free of endotoxins or sepsis risks.
This deep preparation made them ideal partners when collaborating with the academic team that manufactured Sid’s trial vaccine. Rather than acting as passive patients, they approached the scientists as sophisticated collaborators. This mutual technical understanding allowed them to troubleshoot complex regulatory and technical hurdles together in record time—taking the project from initial concept to patient injection in just six months.
- **Knowledge Democratization:** Large language models allow non-clinicians to quickly grasp esoteric jargon, assay formats, and reagent design.
- **First-Principles Quality Control:** AI was used to establish benchmark criteria for custom mRNA purity and sterility.
- **Rapid Execution Timeline:** The personalized mRNA vaccine moved from initial design to human injection in 6 months.
> "What AI has made possible, particularly I think I'm the super user of deep research, is the democratization of that specialized knowledge... it makes me someone who is competent enough to talk to specialists and push some of these programs forward while owning the objective of let's keep Sid alive."
> "AI does feel like an Iron Man suit to allow us to assess all these different specialty things and think about these really arcane, nuanced points... It makes it accessible to somebody like me."
---
[26:54](https://www.youtube.com/watch?v=OAlHiQLsYQM&t=1614s) Rethinking Patient Incentives, Combination Therapies, and Cost
Sid highlights the fundamental misalignment between oncologist incentives and patient priorities. Oncologists are structured to minimize legal and professional liability, making them risk-averse regarding non-standard protocols or severe side effects. Conversely, a patient facing a terminal or aggressive diagnosis wants to maximize overall survivability—even if that means risking severe treatment toxicity. Sid quips that in the face of a brutal cancer, he would far rather risk dying from an aggressive experimental treatment than die from the disease itself.
Most cancer patients pass away because standard-of-care treatments fail sequentially, leaving no time to try alternative therapies once metastases spread. Sid advocates using AI to design aggressive multi-drug combinations run in parallel. When oncologists push back that specific drug combinations haven't been evaluated in $100 million randomized controlled trials (RCTs), patients can use AI to evaluate organ toxicity profiles from first principles—for instance, pairing a drug processed by the liver with one processed by the kidneys to avoid single-organ overload.
Addressing accessibility, Sid emphasizes that fighting cancer aggressively does not always require millions of dollars. While custom biologics are costly, foundational data collection and repurposed generic drugs are surprisingly cheap:
- **Bulk RNA Sequencing:** ~$50 per run.
- **Whole Genome Sequencing:** Starts around $500.
- **AI Research Tools:** $20 per month for frontier-grade models.
- **Repurposed Generics:** Highly affordable if doctors can be persuaded to prescribe them.
> "As a patient, you want to maximize survivability. In my case, because it's a nasty cancer with like medieval medicine, I'd rather die from a treatment than from the cancer."
> "A randomized control trial is a hundred million dollars. We can't test every combination of medicines that way... From first principles, you can just see: do the medicines have the same side effect profile?"
---
[30:01](https://www.youtube.com/watch?v=OAlHiQLsYQM&t=1801s) Scaling Personal Medicine: Building Biotech "On Rails"
To make these advanced workflows available beyond tech founders with deep resources, Sid and Jacob started a portfolio of biotech startups aimed at automating and scaling personalized oncology. Drawing an analogy to Tesla's master plan, Sid notes that while his own approach represents the raw "Roadster Model 001," these companies aim to turn bespoke cancer treatments into accessible "Model 3" offerings using AI as a deflationary engine.
One portfolio company, **Valis**, focuses on maximum diagnostics using single-cell and bulk RNA sequencing to uncover overexpressed targets regardless of tissue origin (such as identifying DLL3 in specific tumors). Valis leverages AI to cross-reference targets against active clinical trials, biological databases, and potential pharma partners. Another company, **Arden**, applies maximum immune profiling to complex, unresolved autoimmune disorders. Arden's CEO uses custom AI tools to instantly digest massive medical histories—such as one patient's 9,000-page medical record—converting unstructured chaos into actionable, step-by-step treatment roadmaps.
Sid updates the audience on his current health status: he remains free of disease following his radioligand therapy and surgery. Through global research, AI analysis, and diagnostic scans, he expanded his personal "therapeutic ladder" from zero viable standard options to over 30 candidate treatments held in reserve.
- **Valis:** Uses RNA sequencing and AI target matching to pinpoint tumor-specific markers (e.g., DLL3) independent of cancer type.
- **Arden:** Applies blood immunoprofiling to autoimmune cases; uses LLMs to parse patient files as large as 9,000 pages.
- **Therapeutic Ladder Expansion:** Advanced global sourcing expanded Sid's viable backup treatment options from 0 to 30.
- **Diagnostic Scans before Drugs:** Advocates for pairing protein-binding scans (e.g., FAP, B7-H3) with radioligand therapies to confirm target presence before administering systemic treatments.
> "If Sid and a few others like him are Roadster Model 001, how do we do the Roadster that somebody can buy off the shelf and work our way from the Model S to the Model 3?"
> "By searching and scouring the world, my therapeutic ladder, the options available to me have gone from zero treatments to 30 treatments that I don't want to use... but it's great to have options."
<FollowUp label="Want to go deeper on how the AND-logic gated CAR-T prevents liver toxicity?" query="Explain how AND-logic gates in CAR-T therapy work to prevent off-target toxicity, using the B7H3 and FAP example from Sid's treatment plan."/>
[34:21](https://www.youtube.com/watch?v=OAlHiQLsYQM&t=2061s) Rethinking Medicine with First Principles and AI
The conversation opens with Sid Sijbrandij explaining that while top-tier medical groups perform excellent work, the healthcare field as a whole has lost its way by relying almost exclusively on randomized controlled trials. He argues that applying first-principles thinking alongside advanced AI could drastically accelerate progress in many complex cases. The host observes that while the public frequently laments the lack of a universal "cure for cancer," Sid's approach suggests that curing cancer will actually happen one patient at a time through personalized medicine. Sid agrees, noting that while the tailored methodologies they are demonstrating might become the standard of care in 30 years, their goal is to bring those techniques into the present for patients who need them today.
The host emphasizes the actionable message for families dealing with a diagnosis: affordable data-gathering tools and accessible AI models now empower patients to analyze complex information, advocate for better options, and improve outcomes. Sid reinforces that patient advocacy is particularly critical for rare cancers, sharing that many survivors attribute their survival to stepping up and taking charge of their medical decisions.
- Sid advocates for combining AI with first-principles reasoning to bypass rigid institutional reliance on randomized controlled trials.
- The future of oncology lies in personalized care rather than a single universal cure.
- Patient self-advocacy is vital for rare diseases; affordable testing and AI tools now make proactive patient involvement feasible.
---
[36:00](https://www.youtube.com/watch?v=OAlHiQLsYQM&t=2160s) The Darkest Moment: A Scared Diagnosis and Misinterpretation
Community member Jason DeLuca asks Sid if he ever hit a point where his situation felt completely hopeless. Sid recalls a terrifying 15-minute ordeal during an executive offsite meeting. While in the middle of a session, he received a message from his general practitioner regarding a recent radiology scan reading: "It's positive, not subtle." Mistakenly thinking "positive" meant good news before remembering medical terminology, Sid opened the report to see his lungs "lit up like a Christmas tree." Because osteosarcoma frequently metastasizes to the lungs, radiology filed a single, dire diagnosis: inoperable, widespread cancer spread.
Sid describes the immediate aftermath: stepping out of the meeting room to process the news for 10 minutes, calling his wife, contacting his CFO and Chief Legal Officer (CLO), and meeting with his board of directors the following day expecting to say goodbye. Out of seven doctors reviewing the file, six signed off on the terminal diagnosis. However, one specialist—Dr. Sant Chawla—questioned the findings, pointing out that osteosarcoma does not typically spread through the lymph nodes. Dr. Chawla estimated a 60% chance it wasn't cancer; further evaluation confirmed the lung lesions were merely remnants of a past COVID-19 infection. Sid reflects on how an AI differential diagnosis tool could have immediately suggested non-cancerous alternatives like COVID-19, saving him from intense panic.
- Sid's lowest moment came when a radiologist misdiagnosed lung inflammation from COVID-19 as inoperable metastasized bone cancer.
- Out of seven specialists, only Dr. Sant Chawla—who has treated over 10,000 sarcoma cases—recognized the abnormal pattern and questioned the reading.
- Sid highlights this incident as a key example where AI could provide vital differential diagnoses to prevent misinterpretations.
---
[38:50](https://www.youtube.com/watch?v=OAlHiQLsYQM&t=2330s) Changing the System and Empowering Patients
Reflecting on how the journey changed his spirit, Sid admits he initially suffered from imposter syndrome, viewing himself as a "delusional tech guy" whose software background wouldn't translate to complex medicine without a PhD. However, he quickly realized that the existing healthcare ecosystem suffers from systemic flaws that software and engineering methodologies can fix. His personal website outlines over ten specific systemic changes needed to make medicine truly patient-first.
The experience inspired Sid to take action on a broader scale. He shares that he hired a full-time staff member, Purnima, to manage patient inquiries coming through his website (**c.com**). The host and speakers discuss how people in biotech and medicine almost universally enter the field for noble reasons, but are hampered by bureaucratic bottlenecks. Rather than fighting bad actors, patient advocates and researchers are battling legacy system complexity.
- Sid established **c.com** to document system-level healthcare improvements and hired a full-time coordinator (Purnima) to handle patient requests.
- The overarching goal is fostering "positive sum energy" across biotech, medicine, and patient groups to rebuild health systems around patient agency.
---
[42:40](https://www.youtube.com/watch?v=OAlHiQLsYQM&t=2560s) Fixing Clinical Trial Bottlenecks and Administrative Friction
Jacob Stern and Sid discuss the massive structural bottlenecks preventing life-saving therapies from reaching patients, pointing to the research of advocate Roxandra on clinical trial abundance. Sid notes that bringing a single new drug to market currently costs over $1 billion, severely restricting medical innovation. He points to Australia's regulatory model, which allows clinical trials to proceed using a simple notification system rather than waiting months or years for pre-approval like the FDA, without compromising patient safety.
Sid also vents about Institutional Review Boards (IRBs) at hospitals, calling them incredibly difficult to deal with due to misaligned incentives. He recounts a personal nightmare where he was wheeled into pre-op at 7:00 AM for surgery, only for the procedure to be delayed until 3:00 PM because the hospital IRB was endlessly debating minor typos, commas, and formatting in consent documents that no one would ever read again. Sid argues for "IRB freedom," which would allow researchers and patients to use independent, certified review boards that operate faster and more affordably.
- **Trial Costs & Regulation**: Bringing a drug to market costs over $1 billion; adopting Australia's streamlined notification model over strict FDA pre-approvals could significantly lower trial barriers.
- **IRB Bureaucracy**: Hospital IRBs create extreme administrative delays over minor paperwork formatting; Sid advocates for "IRB freedom" to allow independent, efficient safety boards.
- **AI Matching**: AI tools can bridge the gap in clinical trial matching by scanning complex electronic health records and diagnostic data to match patients with precision trials.
---
[46:19](https://www.youtube.com/watch?v=OAlHiQLsYQM&t=2779s) Closing Remarks and Upcoming OpenAI Forum Events
As the session wraps up, the host thanks Sid Sijbrandij and Jacob Stern for sharing their journey, emphasizing that their story demonstrates how patient agency and AI can lower technical friction and dismantle legacy complexities. The host reiterates that the OpenAI Forum aims to highlight real-world innovators who leverage AI to reshape industries and save lives.
The event concludes with an announcement of upcoming programming on the OpenAI Forum platform, including a session titled "Reimagining Cultural Heritage with OpenAI," presented in partnership with the Sanskriti Foundation and Asmona. Originally recorded in India, the replay will air on Tuesday, March 24th at 8:00 AM Pacific Time.
- The host thanks Sid and Jacob for demonstrating how AI and patient advocacy can overcome legacy systemic hurdles.
- **Next OpenAI Forum Event**: "Reimagining Cultural Heritage with OpenAI" with the Sanskriti Foundation and Asmona on Tuesday, March 24, at 8:00 AM PT.
## Notable quotes
> [03:05](https://www.youtube.com/watch?v=OAlHiQLsYQM&t=185s) He's like, 'Well, you better start meditating right now because you have super high blood pressure and you might have an aneurysm. And the pain you're feeling might be your aorta starting to burst. So, chill out and also get to the emergency room right now.' — Sid Sijbrandij
> [16:54](https://www.youtube.com/watch?v=OAlHiQLsYQM&t=1014s) T-cells are very killy... basically they shoot out proteins that rip holes in cells that are close to them. — Jacob Stern
> [16:54](https://www.youtube.com/watch?v=OAlHiQLsYQM&t=1014s) The bad news is that my liver was showing up three and a half times as much as the 20 Chinese people they scanned before me. So now we were very worried. If I ever were to take this medicine, I might lose my liver. — Sid Sijbrandij
> [23:38](https://www.youtube.com/watch?v=OAlHiQLsYQM&t=1418s) What AI has made possible, particularly I think I'm the super user of deep research, is the democratization of that specialized knowledge... it makes me someone who is competent enough to talk to specialists and push some of these programs forward while owning the objective of let's keep Sid alive. — Jacob Stern
> [23:38](https://www.youtube.com/watch?v=OAlHiQLsYQM&t=1418s) AI does feel like an Iron Man suit to allow us to assess all these different specialty things and think about these really arcane, nuanced points... It makes it accessible to somebody like me. — Jacob Stern
> [26:54](https://www.youtube.com/watch?v=OAlHiQLsYQM&t=1614s) As a patient, you want to maximize survivability. In my case, because it's a nasty cancer with like medieval medicine, I'd rather die from a treatment than from the cancer. — Sid Sijbrandij
> [26:54](https://www.youtube.com/watch?v=OAlHiQLsYQM&t=1614s) A randomized control trial is a hundred million dollars. We can't test every combination of medicines that way... From first principles, you can just see: do the medicines have the same side effect profile? — Sid Sijbrandij
> [30:01](https://www.youtube.com/watch?v=OAlHiQLsYQM&t=1801s) If Sid and a few others like him are Roadster Model 001, how do we do the Roadster that somebody can buy off the shelf and work our way from the Model S to the Model 3? — Sid Sijbrandij
> [30:01](https://www.youtube.com/watch?v=OAlHiQLsYQM&t=1801s) By searching and scouring the world, my therapeutic ladder, the options available to me have gone from zero treatments to 30 treatments that I don't want to use... but it's great to have options. — Sid Sijbrandij
## Mentioned
OpenAI, OpenAI Forum, Sid Sijbrandij, GitLab, Jacob Stern, Chris Nicholson, Scott McKinney, Y Combinator, osteosarc.com, MDM2, 10x Genomics, ChatGPT, GPT-4, B7-H3, Cole Roybal, FAP, PENXIN3, Tesla, Valis, DLL3, Arden, Purnima, c.com, Dr. Sant Chawla, Jason DeLuca, Roxandra, FDA, Sanskriti Foundation, Asmona
After exhausting standard clinical options for recurrent osteosarcoma, GitLab co-founder Sid Sijbrandij and geneticist Jacob Stern utilized generative AI, first-principles bioengineering, and single-patient regulatory pathways to design and deploy bespoke cancer therapies.
Recurrence, Open Data, and Single-Patient Approvals (3:05)
Osteosarcoma affects fewer than 1,000 Americans annually, and survival following recurrence is typically measured in months under standard treatments. Following a 2022 diagnosis and a subsequent local recurrence two years later, Sijbrandij stepped away from executive routines to manage his care. Operating under the FDA's single-patient Investigational New Drug (IND) pathway—which Sijbrandij noted has a 99.7% to 100% approval rate—the team pursued personalized interventions. To aid outside researchers, Sijbrandij published 25 terabytes of his clinical and genomic data to osteosarc.com.
AI-Driven Bioinformatics and Diagnostic Agent Networks (10:50)
To analyze complex data, Stern constructed multi-agent AI pipelines executing bioinformatic tasks from natural language prompts. When evaluating potential Clonal Hematopoiesis of Indeterminate Potential (CHIP) stemming from chemotherapy exposure, the automated system searched medical literature, identified relevant surface markers, processed 600,000 single-cell data points across longitudinal blood draws, and generated interactive visualization code within 30 minutes to rule out the risk.
Stern and Sijbrandij framed large language models as an equalizer that allows non-specialist researchers to hold rigorous technical dialogues with academic scientists. AI tools enabled the team to evaluate first-principles quality-control benchmarks for bespoke mRNA formulations, compressing the timeline of a personalized neoantigen mRNA cancer vaccine from initial concept to human injection to six months.
Engineered Immunotherapies and Off-Target Safety Gating (16:54)
The team pursued parallelized biological interventions, including custom T-cell receptor (TCR-T) therapies and radiopharmaceutical agents targeting Fibroblast Activation Protein (FAP). In Germany, targeted FAP radioligand therapy achieved 60% necrosis and 20% tumor shrinkage, freeing Sijbrandij's tumor from his dura for surgical extraction. AI data mining also flagged extreme overexpression of the hydrophobic target PENXIN3 (10,000 times healthy tissue levels), which historical water-based assays had missed.
For backup treatments, the team developed a Chimeric Antigen Receptor T-cell (CAR-T) therapy directed at B7-H3. However, diagnostic PET scans in Beijing revealed that Sijbrandij's liver took up 3.5 times more tracer than normal controls, signaling fatal organ toxicity from conventional CAR-Ts. Collaborating with researcher Cole Roybal, they engineered an "AND-logic gated" CAR-T requiring dual recognition of both B7-H3 and FAP before activating cell destruction, shielding healthy liver tissue.
Realigning Clinical Incentives and Scaling Personalized Platforms (26:54)
Sijbrandij criticized systemic misalignments in oncology, arguing that standard practice prioritizes institutional liability over patient survival, leaving patients vulnerable when sequential treatments fail. To balance organ toxicity outside of expensive randomized controlled trials, he advocated assessing multi-drug safety profiles from first principles using affordable tools, including bulk RNA sequencing (~$50 per run), whole genome sequencing (~$500), and repurposed generic medicines.
To translate these methods beyond well-resourced founders, Sijbrandij and Stern initiated startup platforms. Valis applies RNA profiling and AI target cross-referencing independent of tissue origin, while Arden uses language models to extract treatment roadmaps from unstructured medical records up to 9,000 pages long. Through these discovery pipelines, Sijbrandij expanded his candidate therapeutic options from zero to 30 backup treatments, maintaining disease-free status.
Systemic Friction, Misdiagnoses, and Regulatory Reform (36:00)
Sijbrandij highlighted systemic vulnerabilities through personal experience, recalling a moment when a radiologist misread COVID-19 lung inflammation as terminal, inoperable metastases—a consensus backed by six out of seven physicians until an experienced sarcoma specialist identified the anomaly. Sijbrandij noted that diagnostic AI models could flag differential options to prevent catastrophic clinical misinterpretations.
Addressing administrative hurdles, Sijbrandij and Stern argued for structural reforms to reduce the $1 billion cost of traditional drug development. They advocated replacing slow FDA trial pre-approvals with streamlined notification models like Australia's, alongside granting patients "IRB freedom" to use independent review boards to bypass lengthy hospital committee delays over non-substantive documentation edits.
While Sijbrandij remains disease-free with expanded reserve therapies, the broader question remains whether these bespoke, high-intensity methods can transition from expensive proof-of-concept efforts into accessible clinical standards given existing hospital review boards and pharmaceutical regulatory structures.
[3:05] Sid Sijbrandij: He's like, 'Well, you better start meditating right now because you have super high blood pressure and you might have an aneurysm. And the pain you're feeling might be your aorta starting to burst. So, chill out and also get to the emergency room right now.'
[16:54] Jacob Stern: T-cells are very killy... basically they shoot out proteins that rip holes in cells that are close to them.
[16:54] Sid Sijbrandij: The bad news is that my liver was showing up three and a half times as much as the 20 Chinese people they scanned before me. So now we were very worried. If I ever were to take this medicine, I might lose my liver.
[23:38] Jacob Stern: What AI has made possible, particularly I think I'm the super user of deep research, is the democratization of that specialized knowledge... it makes me someone who is competent enough to talk to specialists and push some of these programs forward while owning the objective of let's keep Sid alive.
[23:38] Jacob Stern: AI does feel like an Iron Man suit to allow us to assess all these different specialty things and think about these really arcane, nuanced points... It makes it accessible to somebody like me.
[26:54] Sid Sijbrandij: As a patient, you want to maximize survivability. In my case, because it's a nasty cancer with like medieval medicine, I'd rather die from a treatment than from the cancer.
[26:54] Sid Sijbrandij: A randomized control trial is a hundred million dollars. We can't test every combination of medicines that way... From first principles, you can just see: do the medicines have the same side effect profile?
[30:01] Sid Sijbrandij: If Sid and a few others like him are Roadster Model 001, how do we do the Roadster that somebody can buy off the shelf and work our way from the Model S to the Model 3?
[30:01] Sid Sijbrandij: By searching and scouring the world, my therapeutic ladder, the options available to me have gone from zero treatments to 30 treatments that I don't want to use... but it's great to have options.
Original transcript
[0:09] Good afternoon everyone and welcome to
[0:11] the Open AI Forum. I'm Chris Nicholson
[0:13] and I'm really glad to have all of you
[0:15] joining us today. Um for those who are
[0:18] new, the Open AI Forum is a global
[0:20] community where we bring together people
[0:22] across industries to share how AI is
[0:25] being used in the real world, what we're
[0:27] learning and how we can shape its impact
[0:29] together. Um today's conversation is a
[0:32] powerful example of that impact. Uh
[0:34] we're joined by Sid Sijbrandij,
[0:36] co-founder and executive chair of
[0:38] GitLab, and Jacob Stern, a geneticist
[0:41] working closely with Sid.
[0:43] Together they're going to share a deeply
[0:46] personal and deeply technical story
[0:48] about using AI in the context of a
[0:50] serious illness.
[0:52] I've known Sid for more than a decade
[0:54] now.
[0:55] Uh and I'm very grateful that all of you
[0:58] are here today. Uh so with that, I'll
[1:00] hand it over to Scott McKinney,
[1:01] researcher at Open AI, to more fully
[1:03] introduce Sid and Jacob. Thanks, Chris.
[1:06] I'm Scott McKinney. I'm a researcher at
[1:08] Open AI and I'm thrilled to have Sid and
[1:10] Jacob here today.
[1:11] As Chris mentioned, Sid has been
[1:12] battling osteosarcoma. It's a rare and
[1:15] tenacious form of bone cancer. It
[1:17] affects fewer than a thousand Americans
[1:19] a year. If it recurs after the first
[1:21] line of treatment, survival rates are
[1:22] typically measured in months.
[1:25] I myself have been fighting the same
[1:26] fight since 2020 and I've seen firsthand
[1:28] how sparse, blunt, and frankly gruesome
[1:31] the treatment options are. And I've seen
[1:33] how quickly they can be exhausted.
[1:35] After Sid's cancer came back, faced with
[1:37] the unfaceable, he and Jacob went into
[1:40] founder mode, tackling the problem with
[1:41] intensity and audacity.
[1:43] When the traditional healthcare system
[1:45] brought them to the end of the road,
[1:46] they decided to pave their own.
[1:49] As you'll hear in their talk, they're
[1:50] chaining together bleeding-edge tools in
[1:52] conjunction with AI to push the
[1:54] boundaries of personalized medicine.
[1:56] And what I love about this story is that
[1:57] it's not just about resourcefulness in
[1:59] response to his own medical crisis, but
[2:01] rather as entrepreneurs, they're
[2:03] interested in blazing a trail for others
[2:05] to follow.
[2:06] I was first connected to Jacob and Sid
[2:08] last year in the context of my own
[2:10] diagnosis and they've been enormously
[2:12] generous in sharing their playbook and
[2:13] it's totally changed how I've navigated
[2:15] the situation. Thank you.
[2:18] I like to think that their approach
[2:19] offers a glimpse into the future of
[2:21] medicine that can one day be accessible
[2:23] to all. It's given me hope for my own
[2:25] family and for every family afflicted
[2:27] with cancer.
[2:29] I've been inspired to watch Sid leverage
[2:30] his energy, ambition, and resources to
[2:32] chase his own cure. In my own small way,
[2:35] I feel like I get to do that here at
[2:37] Open AI.
[2:38] AGI may be the tide that lifts all
[2:39] boats, but many of us are particularly
[2:41] excited about how it can transform
[2:43] medicine. And we're working hard to make
[2:45] that vision a reality.
[2:47] The team has created public benchmarks
[2:48] for evaluating models in the healthcare
[2:50] domain, deployed clinical co-pilots in
[2:52] primary care settings, and we're working
[2:54] to democratize access to medical
[2:56] expertise in ChatGPT health.
[2:59] In this regard, I think the future is
[3:00] very bright.
[3:02] So with that, I'm honored to give you
[3:04] Sid and Jacob.
[3:05] Well, well, thank you so much for that
[3:07] wonderful introduction, Scott and and
[3:09] Chris. Uh really excited to be here
[3:12] today and to walk you through my cancer
[3:15] journey and emphasize along the way how
[3:17] AI has helped us work through enormous
[3:20] amounts of data and navigate new
[3:21] treatments.
[3:23] I came to the US more than a decade ago
[3:26] in 20
[3:27] uh 15 with a startup called GitLab. Six
[3:30] years later we took that public.
[3:33] But the next year after going public, a
[3:36] lot happened.
[3:37] You all at Open AI released ChatGPT.
[3:41] Um there was a
[3:42] Nobel Prize awarded to a very
[3:44] interesting technology.
[3:47] But also on the downside, I got some
[3:50] weird feeling when I was doing a bench
[3:52] press. I was doing a bench press and I
[3:54] felt a pain around my heart area and I
[3:57] thought, oh, I know this is not going to
[3:59] be a heart attack because I've had this
[4:01] before.
[4:02] But unlike the few other times, it
[4:04] didn't go away and after 2 weeks I
[4:07] couldn't sleep in the night. At 4:00
[4:08] a.m. I went to um uh I went to the the
[4:13] the hospital, to the emergency room and
[4:15] they they said, "Well, there's no such
[4:16] thing as a 2-week heart attack, so so
[4:19] that's good."
[4:20] And they took an X-ray and sent me home.
[4:21] There was nothing going on according to
[4:23] them.
[4:24] But um
[4:26] uh a few hours later I got a phone call
[4:28] and it was my GP and he said, "Do you
[4:31] know how to meditate?"
[4:32] I thought that was a very vague area
[4:34] question of him and I said, "Yes, I
[4:36] actually I do know how to meditate."
[4:37] He's like, "Well, you better start
[4:40] meditating right now because you have
[4:41] super high blood pressure and you might
[4:44] have an aneurysm.
[4:46] And the pain you're feeling might be
[4:48] your aorta starting to burst. So, chill
[4:51] out and also get to the emergency room
[4:54] right now. I'll call ahead. Um
[4:56] so I did that and they had good news. My
[4:59] aorta was fine, but they also had bad
[5:01] news. There was a 6 cm
[5:04] what turned out to be a tumor growing
[5:06] from my vertebrae.
[5:09] And so I had to
[5:12] do surgery really, really urgently to
[5:15] remove that. They put in a frame, uh
[5:18] they uh did a spine fusion.
[5:21] We did radiation. We did very extensive
[5:23] chemo. And as Scott mentioned earlier,
[5:26] it's kind of medieval how rough that is.
[5:29] It was hard for me to even get to the
[5:32] restroom. I needed four blood
[5:34] transfusions to keep my uh to to even
[5:37] keep alive.
[5:40] As I hinted earlier,
[5:42] 2022, something else happened. It there
[5:44] was a Nobel Prize for click chemistry.
[5:46] You might not have heard of it, but it's
[5:48] a way to combine two compounds in a
[5:51] human without any side effects. It
[5:53] doesn't do anything else. It always
[5:55] happens. It's a very specific reaction.
[5:57] It's genius. That's why it won a Nobel
[6:00] Prize.
[6:01] And I was lucky enough during YC to meet
[6:05] uh someone.
[6:06] Jose is his name and he was starting a
[6:09] company around it. He was first in
[6:10] patients with click chemistry. And over
[6:14] time, he struggled to fundraise. Now
[6:17] he's okay, but at that time, before the
[6:19] Nobel Prize, it wasn't trendy. Biotech
[6:21] VCs didn't believe in it and over time I
[6:23] became his biggest investor. It became
[6:26] my biggest investment. And we ended up
[6:28] becoming best friends.
[6:30] And now that I was in trouble, he said,
[6:33] "You know what? Let's try this treatment
[6:35] for you."
[6:36] And that opened up my eyes that it was
[6:38] possible to get a treatment just for a
[6:40] single patient. I always thought it had
[6:42] to be a big trial. But if you're in a
[6:44] really bad spot, the uh there's a an
[6:48] option called a single patient IND.
[6:50] The FDA approves 99.7%.
[6:54] Uh they even say now that they're
[6:55] approving 100% and it's a great pathway
[6:58] if you're uh
[7:00] with a disease and and there are very
[7:02] few options.
[7:04] Despite all of our efforts, 2 years
[7:07] later, we had a remission. It started
[7:09] growing again. It had a local
[7:11] progression. And there were no more
[7:13] standard of care medicines. My
[7:15] oncologist couldn't recommend something
[7:17] that he thought was going to make a
[7:18] difference. He said, "Hey, go look for
[7:21] trials." But it's a rare disease. There
[7:23] were no trials.
[7:25] So this became
[7:27] life or death for me and I quit my day
[7:30] job and started going founder mode on my
[7:32] cancer.
[7:34] And I went all out. I did all the
[7:37] diagnostics I could could my get my
[7:40] hands on.
[7:41] Typically people say, "Only do the
[7:43] diagnostics if you know like what you're
[7:45] going to do with it." We didn't do that
[7:47] at all. We did every single technology
[7:49] we could find and and uh we collected a
[7:52] lot of data.
[7:53] We started making treatments. If there's
[7:55] no treatment left, you need to make
[7:57] them.
[7:58] And we started doing as much as possible
[8:00] in parallel. We were out of time and
[8:03] most people die because
[8:05] the disease progresses.
[8:07] And we're trying to scale it for other
[8:09] people.
[8:11] In the maximal diagnostics, we've done
[8:13] everything under the sun and if you want
[8:15] to see some of that, there's 25
[8:16] terabytes of data available on
[8:18] osteosarc.com.
[8:21] One of the things we did was single-cell
[8:23] sequencing. It's a really cool
[8:25] technology where you individually sample
[8:27] thousands of cells. It allowed us to
[8:29] isolate my cancer cells and look at the
[8:32] properties of the cancer cells. And we
[8:34] found out they had a lot of FAP. It's
[8:37] kind of a fibrous tissue.
[8:39] And that was great because we found in
[8:41] Germany
[8:43] a doctor that had an experimental
[8:45] treatment where you combine FAP
[8:48] with radioactive substances. So I went
[8:51] there, did it twice. I was in iso- the
[8:53] isolation ward afterwards because you
[8:55] were kind of you had to glow up. Um
[8:58] and it was really, really successful.
[9:01] The two treatments led to 60% necrosis
[9:04] and 20% shrinkage. And the cancer
[9:06] detached from my dura and we were able
[9:08] to go in and scoop it out of my body.
[9:11] And when we looked at the sequencing
[9:13] we've done along the way, the TCR
[9:15] sequencing in particular, we saw that
[9:18] we've been able to all the immuno three
[9:21] treatments
[9:22] get a lot of angry T cells in me. So I
[9:25] did everything under the sun. I did dual
[9:27] checkpoint inhibitors, NK cells, a super
[9:31] antagonist, an oncolytic virus. So we
[9:33] had my immune system riled up. But the
[9:36] tumor microenvironment was so was
[9:39] suppressive. It didn't allow the immune
[9:41] system to go in. And we suspect what
[9:43] happened is that because we disabled the
[9:45] FAP cells, they stopped signaling to the
[9:48] neutrophils and we turned the cold tumor
[9:51] hot. Suddenly the immune cells were able
[9:53] to go in. We don't know this for
[9:54] certain, but that's what we think
[9:56] happened.
[9:59] We also looked at MDM2 as a
[10:02] therapy for me.
[10:03] I am really, really high on the MDM2
[10:06] expression. It's just some protein, but
[10:08] I'm I'm off the scale. I'm the index
[10:10] tumor over here. So, I'm on the 0.3%
[10:14] and we started looking like, are there
[10:15] any drugs? And there were drugs. People
[10:17] made them. But then they stopped all the
[10:20] development because drugs only get to
[10:22] market if it's a blockbuster.
[10:25] If
[10:26] the majority of patients are served with
[10:28] it. That wasn't the case for MDM2.
[10:31] And so they stopped it and they were
[10:32] about to turn off the freezers. So,
[10:34] we're now paying to keep the freezers on
[10:36] and I'm looking for a pathway to bring
[10:38] this medicine to market because it might
[10:40] help other people beside me. Not the
[10:42] majority of patients, but if there's 10
[10:44] or 20% that'd be really awesome outcome
[10:47] as well.
[10:50] One of the best things that happened
[10:51] with Maximal Diagnostics is that the
[10:53] single cell sequencing and got me in
[10:55] touch with Jacob, who's now running the
[10:57] enterprise of my care. Jacob. Go.
[11:00] Thanks, Sid. Um and thanks so much for
[11:02] having us. Um you know, I'm not a
[11:05] doctor. Uh I met Sid through Jose, the
[11:07] mutual friend, who uh was working on the
[11:10] click chemistry-based drugs.
[11:12] And at the time I was working at 10x
[11:14] Genomics, a company that makes the
[11:16] equipment that enables a single cell
[11:17] sequencing. I was there for 6 years.
[11:20] And when I met Sid, I started talking
[11:22] about what we're doing at my 10x. We do
[11:23] single cell.
[11:25] Sid explained how he and his team were
[11:26] using single cell to actually inform his
[11:28] care.
[11:29] And at 10x, the mission statement is to
[11:32] master biology to advance human health.
[11:34] And this was the first time I'd actually
[11:35] met someone in person who was doing
[11:37] this. He was mastering biology to
[11:38] advance his own human health.
[11:40] Uh and I found this to be super
[11:42] compelling and as we got to know each
[11:44] other and I got to know the story, uh I
[11:46] got to understand the extent to which he
[11:48] was living in the future and how to
[11:50] envision to do even more.
[11:51] And so, what you can see here is the
[11:54] vast array of of technologies we're
[11:56] using to analyze Sid's cancer. We're
[11:58] doing stuff that I'm more familiar with
[12:00] like single cell sequencing uh or some
[12:02] of the bulk DNA or RNA sequencing, but
[12:05] we're also stretching way outside of my
[12:07] comfort zone. We're doing targeted radio
[12:09] diagnostics, extensive pathology
[12:11] staining, using organoid models to test
[12:13] the effect of drugs directly on Sid's
[12:15] cancer.
[12:16] And so, for me,
[12:19] I've become an extensive user of AI to
[12:21] basically bring myself up to par
[12:23] and uh enable me to talk with experts
[12:26] about these different domains and have
[12:28] intelligent conversations about how we
[12:30] can drive this forward for Sid and
[12:32] actually make use of this data. So, I'm
[12:34] actually going to take you through a
[12:35] little bit of my historical chat GPT
[12:37] history so you can see a few examples of
[12:39] how I actually did this in practice.
[12:41] Um so, this is a a screen a screenshot
[12:44] from last summer where we wanted to just
[12:46] run an experiment. Let's take the output
[12:49] of one of these bulk RNA sequencing
[12:50] experiments where we're basically
[12:52] counting the number of times that each
[12:55] gene is being is being detected at the
[12:57] RNA level in the sample from Sid's
[12:59] tumor. Uh and this is expressed in a CSV
[13:02] file. All right, it's genes and then
[13:03] counts.
[13:04] And we fed this to at this point it was
[13:07] uh 400 on the pro plan and said, "What
[13:09] do you think of it?" And I wanted to see
[13:11] what came out. And frankly, even then it
[13:13] was remarkable. Uh you'll see here if
[13:16] you zoom in, uh it it flagged B7H3,
[13:18] which you'll recognize later in the
[13:20] presentation. And then it also
[13:21] recognized some of the immune dynamics
[13:24] uh that Sid talked about earlier, which
[13:26] we've studied in much more detail since
[13:28] uh at the single cell level.
[13:31] And now what we're doing is much more
[13:33] advanced. I mean, these models are
[13:34] progressing at an absolutely insane
[13:36] rate. Uh and we've built some harnessing
[13:38] for ourselves where we can ask a we can
[13:41] ask a question in natural language and
[13:43] spin up a series of agents that can go
[13:46] uh do literature search, formulate its
[13:48] own hypothesis, basically structure uh
[13:51] structure a bioinformatic analysis and
[13:54] then execute that and bring back a
[13:55] summary.
[13:56] So, what I have here is a recent example
[13:59] where I'm asking this system that we
[14:00] built about chip clonal hematopoiesis of
[14:04] indeterminate potential. This is
[14:06] something that many people get as they
[14:08] age. Uh it's basically the blood the
[14:10] blood cells losing polyclonality as we
[14:12] age uh and it can, if things go wrong,
[14:15] lead to leukemia. And as we got sort of
[14:17] a hint that this might be happening in
[14:19] Sid from our uh from our team, it
[14:21] spooked us because this can be a side
[14:23] effect of the chemotherapy that Sid took
[14:25] a number of years ago.
[14:27] And so, one of the first things I did is
[14:29] I went to the system and just asked. You
[14:30] can see the the question that I posed
[14:32] here. It's pretty natural in API costs.
[14:34] It thought for 30 minutes.
[14:39] Uh did a series of tool calls. It did
[14:41] its own literature review, formulated
[14:43] the set of markers it wanted to look
[14:45] for. This is hooked up to about 600
[14:48] 600,000 single cells from uh a number of
[14:51] time points that we've taken from Sid's
[14:52] blood. So, it can actually run the
[14:54] analysis directly in Sid's blood and
[14:57] come back to me with a report of here's
[14:59] what I think, here's my conclusions,
[15:02] here's the interactive plots, and then
[15:04] here's my whole history, here's the code
[15:06] it wrote in Python, etc.
[15:08] And I'm not going to trust this out of
[15:09] the box. Um it can certainly make
[15:11] mistakes, but what it does give me is
[15:14] it's a rapid way uh for me to get up to
[15:16] speed and start to understand
[15:19] the sort of circumstances around this
[15:20] disease. Again, I'm not a doctor. I
[15:22] didn't train in chip. But here, I can
[15:24] rapidly come up to speed on the disease
[15:26] more generally and also on Sid's biology
[15:28] specifically. And it's led us it's led
[15:31] me be a good counterpart to the
[15:32] informaticians who have since looked
[15:34] into this in much more depth and
[15:36] thankfully we put this risk to bed.
[15:38] Um
[15:39] so, on top of doing the maximum
[15:41] diagnostics, we're also uh making a
[15:44] number of treatments from scratch
[15:46] specifically for Sid. And uh you know,
[15:48] this is even further outside of my
[15:50] comfort zone than the analysis is and
[15:52] it's frankly been radicalizing for me to
[15:53] find that you can just make your own
[15:55] drug. And so, before Sid goes through
[15:58] and talks about some specific vignettes
[15:59] here, I just want to sort of give an
[16:01] overview of some of the modalities uh
[16:04] just to just to level set. So, the first
[16:06] one he'll talk about is the cancer
[16:08] vaccine. Uh in this case, an mRNA
[16:10] vaccine specifically.
[16:12] So, you can think about this similar to
[16:14] the vaccine that you've taken for COVID
[16:16] or for the flu, whatever it is, where
[16:18] the idea here is to educate the immune
[16:20] system to prepare itself to fight
[16:23] something that's foreign.
[16:24] Uh in the case of the COVID vaccine,
[16:26] you're exposing the body to the spike
[16:28] protein of of of COVID itself. In this
[16:31] case, we've encoded a number of
[16:33] mutations that are present in Sid's
[16:35] cancer, which basically differentiate
[16:37] the foreign cancer from Sid's normal
[16:39] tissue into the vaccine. And when we put
[16:42] that in, we're priming his immune cells
[16:44] uh to be ready to patrol and try and
[16:46] fight the cancer if it sees it.
[16:48] Uh the next vignette will be around a
[16:50] TCR T cell therapy. Uh TCR stands for T
[16:54] cell receptor. And T cells are one of
[16:56] the business ends of the immune system.
[16:59] Uh T cells are very killy. They uh they
[17:01] use a T cell receptor to detect
[17:03] something very specific and then uh
[17:06] basically they
[17:07] shoot out proteins that rip holes in
[17:09] cells that are close to them.
[17:11] And so, what we can do here is something
[17:12] that's a little bit more direct than
[17:14] what we're doing in vaccines,
[17:15] where through some of the work we've
[17:17] done with single cell sequencing and uh
[17:20] some very capable partners have done
[17:21] very thorough work, as you'll see, uh
[17:24] we've identified TCRs T cell receptors
[17:26] that are likely very specific for Sid's
[17:28] cancer. And we can engineer those
[17:30] directly into healthy T cells, grow a
[17:32] bunch of those up, and then inject them
[17:35] directly into Sid. Hopefully, we never
[17:36] have to do that, but we're preparing in
[17:38] case we need to.
[17:39] Uh the last one I'll talk about briefly
[17:41] is the CAR T cell therapy. In this case,
[17:44] CAR stands for chimeric antigen uh
[17:46] receptor.
[17:48] This is an extension of that same
[17:49] concept of the TCRT,
[17:51] where we're taking a T cell, the killer
[17:53] immune cell, as a chassis
[17:55] and souping it up and then engineering
[17:58] in the receptor, basically what it's
[18:00] going to use to recognize. And we picked
[18:02] a target that through the maximum
[18:04] diagnostics, we saw as present on, you
[18:07] know, most if not all of Sid's cancer
[18:09] cells.
[18:10] And through this, this is this is the
[18:11] really killy one, where if we really
[18:13] needed a nuke to deploy, we're putting
[18:15] this on the shelf as an insurance
[18:16] policy, a sort of super killy CAR T
[18:19] that's specific and very potent.
[18:24] Yeah, so we did the personalized mRNA
[18:27] vaccine. It's super exciting. I was the
[18:29] first patient of an
[18:30] investigator-initiated trial.
[18:33] And while doing that, it was awesome to
[18:36] learn what goes into making it. There's
[18:38] all these antigens that you you can
[18:41] encode in it. You have
[18:43] bits and bytes basically to encode base
[18:46] pairs to encode in the virus. You have a
[18:47] limited set. So, what you select
[18:50] matters. And we used tons of different
[18:52] ways to determine that. Like, what was
[18:54] in my tumor samples? What is scoring
[18:57] high in different models?
[18:59] And right now, this is more of an art
[19:01] than a science, but we've already seen
[19:03] companies that are starting to use AI to
[19:06] do this automatically. And that is very,
[19:09] very exciting because that means you can
[19:11] do it for millions of patients. I can
[19:13] see a future where
[19:15] you get a personalized vaccine against
[19:19] your cancer. And that will have to come
[19:21] from AI because we don't have enough
[19:23] doctors to do this. Also, typically
[19:26] these engineering AI approaches yield
[19:29] better results in our opinion. So, very,
[19:32] very exciting what AI can do here.
[19:35] We were very fortunate that we got this
[19:38] from project start to injection in only
[19:40] 6 months. And
[19:42] we see those times coming down even
[19:44] further in the future.
[19:46] Jacob mentioned the TCRT.
[19:49] It's a very interesting
[19:51] thing to make as It's very complex to
[19:54] make,
[19:55] but a lot of decision-making goes into
[19:57] it. A lot of data goes into it. And what
[19:59] we've seen is that the diagnostics,
[20:02] if you have good kind of tumor samples
[20:05] and you do a lot of sampling, it really
[20:07] helps inform how you design it. Here
[20:10] again, there were lots of decision
[20:12] variables like do we do that? Yes, no,
[20:15] why? A lot of reasoning, a lot of logic.
[20:18] And to do this at scale, you'll need AI
[20:21] as well.
[20:22] The last example I'll give is, I think,
[20:25] a really powerful one of how the
[20:28] diagnostics are influencing the medicine
[20:31] development.
[20:32] We selected B7H3
[20:35] as a target for my CAR-T. The CAR-T,
[20:38] remember that's that super killing
[20:40] nuclear bomb that you're going to set
[20:42] off.
[20:43] But we wanted to know where is that
[20:45] present? And you have genetic data about
[20:47] me, but there's nothing as good as doing
[20:50] an actual scan. And we were able to
[20:52] access one. It was in China, in Beijing.
[20:55] So I traveled there in October. And we
[20:57] did the scan. And the doctors came back
[20:59] and he said, "We have really good news
[21:01] and really bad news."
[21:03] Well, what's the good news? The good
[21:05] news is they didn't see any cancer. So
[21:07] yay, still clean.
[21:09] The bad news is that my liver was
[21:13] showing up three and a half times as
[21:15] much as the 20 Chinese people they
[21:18] scanned before me.
[21:19] So now we were very worried. If I ever
[21:22] were to take this medicine, I might lose
[21:25] my liver.
[21:27] So we went back to Cole Robo who's
[21:30] making it. He said, "Can you do
[21:32] something about it?" He says, "Well,
[21:34] actually, I was the first one to invent
[21:36] a larger gate for CAR-Ts, an end gate.
[21:39] Two things have to be present." So we
[21:41] went back and looked at like what's
[21:43] really not present in the liver. It's
[21:45] FAP. You know that from the first one,
[21:47] the radio ligand therapy. That has very
[21:50] low liver expression.
[21:52] And by needing those both things to be
[21:55] present, the B7H3
[21:57] and the FAP, we're going to make sure
[22:00] that if I need use the nuclear bomb, it
[22:02] doesn't go off in my liver.
[22:04] And it was amazing to see that feedback
[22:06] loop, how a better scan informed that.
[22:10] Last anecdote is something we saw. It's
[22:13] called PENX3.
[22:14] And it it it it
[22:16] on the genetic data, it like sprung at
[22:19] us. It was represented
[22:21] 10 times 10,000 times as much in my
[22:25] cancer as in my healthy tissue.
[22:28] That's that's an incredible difference.
[22:31] And we were eager like, "Okay, let's see
[22:33] what medicines are available." But when
[22:35] we scoured the literature, there was
[22:36] nothing. There there were a few a few
[22:38] mentions of it. No one's really done the
[22:40] research. And most protein targets,
[22:43] people will have published on it. They
[22:45] they they just go look for what's
[22:47] expressed on the outside of a cell. So
[22:48] this was really strange. Why wasn't
[22:50] there anything?
[22:51] And we think we figured it out. It's a
[22:54] hydrophobic thing. So if you do your
[22:56] tests in water, like most people do,
[22:59] you're not going to find this thing. The
[23:01] only reason we found it is because we
[23:03] went back to the data and did a
[23:05] painstaking analysis. And this is again
[23:08] something that AI is incredible for. It
[23:10] has much more patience than a human. It
[23:12] will go through all those
[23:15] terabytes of data and find the proteins
[23:18] that matter. A human would not have the
[23:20] patience for this. And we were
[23:22] lucky that we we were able to find this.
[23:25] And now we're running a binder campaign
[23:26] trying to see if we can engineer
[23:28] something against this.
[23:31] So that's some of the anecdotes of how
[23:34] we made the personalized drugs and
[23:36] diagnostics. Back to you, Jacob. Yeah.
[23:38] So
[23:40] uh once again, we'll dive back into my
[23:41] ChatGPT history. Um
[23:44] I would say this is an area again where
[23:46] I'm even more out of my depth than I am
[23:47] with some of the more adventurous uh
[23:50] analysis modalities. And so what I'm
[23:52] doing absolutely would not have been
[23:53] possible 2 years ago. Because if I were
[23:55] trying to uh navigate to the expert
[23:58] who's at some biotech company, you know,
[24:02] it advancing, you know, some program
[24:04] that's going to be used for millions of
[24:05] people, just the knowledge is very
[24:07] sparse. But what AI has made possible,
[24:10] particularly I think I'm the super user
[24:11] of deep research, is the democratization
[24:14] of that specialized knowledge. And it
[24:16] doesn't make me a specialist, but it
[24:18] makes me someone who is competent enough
[24:20] to talk to specialists and push some of
[24:23] these programs forward while owning the
[24:25] objective of let's keep Sid alive.
[24:28] And so here you can see some of the
[24:30] stuff we were looking at as we were
[24:31] thinking about the mRNA vaccine. Uh last
[24:34] summer as we were considering how are we
[24:36] going to do this, we were looking at
[24:37] options ranging from working with one of
[24:40] these, you know, big professional
[24:42] contract development manufacturing
[24:44] operations that makes drugs
[24:46] professionally, probably makes the COVID
[24:48] vaccine actually, uh down to could we
[24:51] actually just have some academic group
[24:53] make it in their lab as if they were
[24:54] putting it in a mouse and would we put
[24:56] that in Sid? And so for that, we
[24:58] actually thought from first principles
[25:00] around what are the quality control
[25:01] metrics we would want to see from
[25:03] something in order to feel comfortable
[25:05] that what we're putting in Sid is going
[25:07] to work, it's going to be what we
[25:08] expect, it's going to be clean, it's not
[25:10] going to give him sepsis, etc. And of
[25:12] course all that work was done with AI.
[25:15] Um
[25:16] and that meant that when we were
[25:17] introduced to the academic team that
[25:19] ended up making the mRNA vaccine, which
[25:21] has been an amazing collaboration, uh it
[25:24] did two things. One, we were prepared to
[25:27] ask reasonable questions and understand
[25:29] what they were saying and what they were
[25:30] planning to do. And perhaps even more
[25:32] importantly, it made us good partners to
[25:33] them. Uh because we were sophisticated
[25:36] enough to understand what they were
[25:38] doing. And that built a relationship of
[25:40] trust. And we had to fight through a lot
[25:43] of things to actually get this done. And
[25:44] we did that together. And that was built
[25:46] on sort of this mutual understanding and
[25:48] mutual ability to talk about the
[25:50] nitty-gritty details here.
[25:52] Uh
[25:53] here this is on PENXIN3. As Sid
[25:55] explained, this is a really understudied
[25:57] target. When we're thinking from first
[25:59] principles about Sid, if you go back to
[26:01] that plot, uh we don't have to go back
[26:03] there, but on that plot, the PENXIN3 was
[26:06] way off the diagonal. It was clearly the
[26:09] best target for Sid.
[26:11] Um but it's not the best target for
[26:13] almost anybody else. And so it's a
[26:15] target that's understudied. And so as
[26:16] we're trying to do a binder discovery
[26:18] campaign against PENXIN3, uh we're sort
[26:21] of MacGyvering through it. We're
[26:23] inventing new reagents. We're thinking
[26:25] through new assays. Uh we're trying to
[26:27] triage different data sites to see if we
[26:30] have a good binder or not. Is it a false
[26:32] positive or a false negative? And as
[26:34] we're thinking through that, like AI
[26:36] does feel like an Iron Man suit to allow
[26:38] us to assess all these different
[26:40] specialty things and think about think
[26:43] about these really arcane nuanced uh
[26:46] points in this like arcane process. And
[26:48] it makes it accessible to somebody like
[26:50] me.
[26:54] If we think about kind of
[26:57] the treatments in parallel, what where
[26:59] did that come from?
[27:01] I think something you have to be aware
[27:03] of as a patient is that the incentives
[27:05] for the doctors are very, very different
[27:08] from the incentives for you as a
[27:10] patient.
[27:11] The doctors want to minimize their
[27:14] liability. And their liability is when
[27:16] they prescribe a treatment and
[27:19] there is severe side effects and
[27:21] including death from those effects.
[27:24] As a patient, you want to maximize
[27:26] survivability.
[27:29] In my case, because it's a nasty cancer
[27:32] with like medieval medicine, I'd rather
[27:35] die from a treatment than from the
[27:37] cancer. Dying from cancer is a really
[27:39] miserable way to go.
[27:41] That is not what you will get with
[27:44] doctors. They have very different
[27:46] incentives than you.
[27:48] And most patients, they get too few
[27:52] medicines and they die because the
[27:54] medicines that they got didn't work.
[27:56] They didn't have time to try the other
[27:58] things. They get metastatic cancer and
[28:00] they pass away.
[28:03] AI is amazing at helping you suggest
[28:06] things to discuss with your oncologist
[28:11] to combine things.
[28:13] And the standard, I think, pushback you
[28:16] get from an oncologist is this has never
[28:18] been tested in a randomized control
[28:21] trial.
[28:22] That is true. But a randomized control
[28:24] trial is a hundred million dollars. We
[28:27] can't test every combination of
[28:29] medicines that way. It also it's not
[28:32] needed. From first principles, you can
[28:34] just see do the side effects do the
[28:37] medicines have the same side effect
[28:38] profile? Do they add up? You almost
[28:41] would you kind of need to look at this
[28:43] organ by organ. If one drug is hitting
[28:46] your kidneys, you want the other drug to
[28:48] hit your liver. You don't want to take
[28:49] two kidney drugs at the same time. So AI
[28:53] is incredibly useful to help you
[28:57] through that, to have that dialogue with
[28:59] your doctor, but to come well informed
[29:01] and to push them to get closer to
[29:03] maximum survivability because that will
[29:06] not be their default mode.
[29:09] If you want to see what treatments I've
[29:12] combined, um you can see see so on
[29:15] osteosarc.com/timeline
[29:17] where we have all my different
[29:18] treatments and all the modalities listed
[29:21] out.
[29:23] Some of the things I'm doing are
[29:25] incredibly expensive.
[29:27] Some of the things I'm doing are not
[29:29] incredibly expensive. You can get bulk
[29:32] RNA sequencing these days for 50 bucks.
[29:35] Whole genome sequencing started $500.
[29:39] AI is amazing. For $20, you get super
[29:42] capable tools.
[29:44] Treatments in parallel will consume more
[29:46] drugs. but if those drugs are generics,
[29:49] the the pricing is super affordable.
[29:52] So, there's there are things that should
[29:56] be accessible to people if they can
[29:58] convince their doctor to prescribe them.
[30:01] And then aside from that, for some of
[30:04] the more adventurous things that we're
[30:05] doing,
[30:06] part of the way we're trying to scale
[30:07] this for others is we've started a
[30:09] series of companies to try and put this
[30:11] on rails. Cuz for Sid and a few other
[30:13] families who are sort of pursuing this
[30:16] at the absolute maximum, it's been the
[30:18] coming up the learning curve and
[30:20] starting from scratch and developing
[30:22] relationships and developing
[30:23] competencies, etc. But what we want to
[30:26] do is sort of pave the road behind us so
[30:28] that it's easier for the next person,
[30:30] the next person, the next person. If Sid
[30:32] and a few others like him are Roadster
[30:34] Model 001, how do we do the Roadster
[30:36] that somebody can buy off the shelf and
[30:38] work our way from the Model S to the
[30:40] Model 3? Uh and of course AI is deeply
[30:43] embedded in all of these things because
[30:45] it's a deflationary technology. It's how
[30:47] we get things It's how we bend the cost
[30:49] curve and how we get things to scale.
[30:51] And so, I'll give a couple of examples
[30:53] through the portfolio. Uh one of the
[30:55] companies is Valis.
[30:57] Uh they're doing the maximum diagnostics
[31:00] from a uh gene expression perspective,
[31:03] running things like single cell
[31:04] sequencing and running through all the
[31:06] raw bulk RNA sequencing data to find the
[31:09] right target for somebody's cancer
[31:11] regardless of the tissue of origin.
[31:14] And so, here you can see sort of the
[31:16] plot looking at a specific patient and
[31:19] the expression levels, etc. But once the
[31:21] uh once the target's identified, in this
[31:23] case DLL3, using AI to pull in all sorts
[31:26] of context about this target, active
[31:28] clinical trials, biological data,
[31:30] pharmaceutical companies to work with,
[31:32] etc.
[31:34] And then
[31:35] for another one, Arden.
[31:37] Uh Arden is taking a similar approach to
[31:39] complex unresolved immune disorders by
[31:41] doing maximum profiling of blood uh to
[31:44] try and get into a treat, measure,
[31:46] analyze, repeat cycle with combination
[31:49] of targeted uh
[31:50] targeted immunomodulators that are
[31:53] tailored to whatever the person's blood
[31:54] is actually saying.
[31:55] And the the CEO of Arden, when he starts
[31:58] working with somebody, often what comes
[32:00] back is a gigantic Google Drive folder
[32:02] or in one case a 9,000-page document
[32:05] documenting somebody's medical history.
[32:07] And so, he's actually used AI to build
[32:09] an AI tool that helps him parcel this
[32:12] information and get up to speed with the
[32:15] patient and help them uh parse the
[32:17] history to make a better plan going
[32:18] forward.
[32:21] So,
[32:22] I've been very, very lucky that what we
[32:25] did worked and there's no evidence of
[32:27] disease since we did the radioactive
[32:29] treatment and uh surgery. It's not for a
[32:32] lack of trying.
[32:34] Uh couple of weeks ago, I did a bunch of
[32:36] experimental scans. They These scans
[32:39] combine a protein binder, for example,
[32:42] for B7H3, but also FAP2, FAP with a
[32:47] scan. And I'm super excited that there's
[32:49] more and more of these proteins kind of
[32:51] becoming available so you can do a scan.
[32:53] I can see a future where instead of just
[32:55] trying a drug,
[32:57] even a generic drug, if there's a
[32:59] protein it binds to, you want to make
[33:01] sure that is expressed in your cancer
[33:04] and you want to be aware where else it's
[33:06] expressed so you can view for those side
[33:08] effects. That differs per person. If we
[33:10] can do the scan first, that's great.
[33:12] With some of these, you can do the scan
[33:14] first and then you can do the scan in
[33:16] the morning and you can do the treatment
[33:17] in the afternoon where they bind
[33:19] something radioactive to it. I think
[33:21] this field is
[33:22] poised to take off really, really
[33:24] rapidly.
[33:26] By searching and scouring the world, my
[33:28] therapeutic ladder, the options
[33:30] available to me have gone from zero
[33:32] treatments to 30 treatments that I don't
[33:35] want to use, I'd rather not use, but
[33:37] it's great to have options.
[33:40] I want to thank you so much for watching
[33:43] us and we super look forward to your
[33:44] questions. Thank you.
[33:48] Hey guys.
[33:49] That was amazing. Thank you so much.
[33:51] Yeah. Um
[33:52] one of my takeaways that that I'm not
[33:55] you know, people maybe who haven't
[33:57] suffered cancer or haven't gone through
[33:59] cancer like you, they may be surprised
[34:00] is how dynamic and proactive
[34:05] uh you have been with cancer. You've
[34:07] you've learned from it so quickly. And
[34:10] it feels like that the speed of your
[34:12] learning process is kind of the key to
[34:14] getting this far. Is that how you feel?
[34:17] Yeah, and we we learned a lot from other
[34:18] people, other patients, uh concierge
[34:21] medicine groups that are super, super
[34:23] good, but also a lot of AI and a lot of
[34:26] first principles. I think medicine has
[34:29] lost its way in a little bit that
[34:31] the
[34:32] only work with randomized controlled
[34:34] trials, but there's some first
[34:36] principles thinking would would really
[34:38] help in a lot of lot of cases. Yeah. Um
[34:41] people people throw around the term a
[34:43] cure for cancer a lot and lament that AI
[34:46] has not given us the promised cure for
[34:48] cancer. But like I I feel like what I'm
[34:51] learning from you is that
[34:52] curing cancer may happen one patient at
[34:55] a time.
[34:57] Right? So, you're really painting a
[34:58] future of personalized medicine where
[35:00] there's not a single cure.
[35:03] Yeah, it's we're trying to bring that
[35:06] I think some of the stuff we presented
[35:08] today is going to be the standard of
[35:10] care 30 years from now. We're trying to
[35:13] bring it to the present. Some of these
[35:16] to the present for older people
[35:19] and all of this to the present for some
[35:21] people.
[35:22] Yeah. I I heard you making some
[35:23] recommendations. I know there's a lot of
[35:25] families out there who have had brushes
[35:28] with cancer or who are going through
[35:29] cancer now. And what I hear you saying
[35:31] is there's affordable ways to gather
[35:34] data and there are affordable tools help
[35:36] you analyze and understand that data so
[35:38] that you you may ask for better
[35:40] treatments and eventually get to better
[35:42] outcomes. Is that Is that the gist of
[35:44] your message? Especially if you have a
[35:46] rare cancer, it's super important to be
[35:48] a good advocate. And even if you have a
[35:50] known rare cancer, we've met many
[35:52] survivors who said that being being
[35:54] stepping up their advocacy has really
[35:56] paid off for them. Yeah, amazing. Okay,
[35:58] I'm going to open this up to um
[36:00] questions from the community. So, we got
[36:02] Jason DeLuca who asks, "Did you ever hit
[36:06] a point where things felt absolutely
[36:08] hopeless? And what was your breakthrough
[36:10] moment with the tech?"
[36:13] I think um
[36:16] it wasn't a tech thing, but I felt very,
[36:18] very hopeless for 15 minutes when I
[36:22] um
[36:23] I got uh radiology result.
[36:27] And it was a classic. I was in a
[36:29] meeting, but all of my reports um it was
[36:32] um during an offsite and I got a message
[36:35] from my GP. He says,
[36:37] "Uh
[36:38] it's positive, not subtle."
[36:40] Oh, okay, great, positive. Uh
[36:43] and then I'm like, "Oh, no, no, wait.
[36:44] It's the the reverse in medicine. What's
[36:46] going on?" So, I opened it. My lungs lit
[36:49] up like a Christmas tree and um
[36:52] with bone cancer, you're really afraid
[36:54] of spreading to the lungs.
[36:56] And it was a single diagnosis from
[36:59] uh from the
[37:01] uh from radiology that the cancer has
[37:03] spread and uh there were so many it was
[37:06] inoperable.
[37:08] And so, that was I walked out of the
[37:11] room uh
[37:13] uh you process for 10 minutes, you call
[37:15] your wife, you call your CFO and CLO and
[37:19] then you go from there. And the next
[37:21] day, the board came over. That was That
[37:23] was fun.
[37:24] Um and you're kind of like, well, this
[37:26] been a ride. Thanks, everyone. All the
[37:28] doctors signed off like,
[37:31] "Oh, so sorry to hear about the news."
[37:33] And then of the seven doctors we had in
[37:35] the loop,
[37:36] one guy, one guy like, "I don't like how
[37:40] this spread by the lymph nodes. This
[37:42] This is not how osteosarcoma spreads
[37:45] through the lymph node."
[37:47] And he's like, "Yeah, 60% this isn't
[37:50] cancer."
[37:52] And it turns out it was the remnants of
[37:54] COVID.
[37:55] Uh but that was certainly a low point.
[37:58] Yeah. But yeah, you also learn 10
[38:01] minutes plus 5 minutes. Yeah. Uh and
[38:04] you're you're back and you're you're
[38:05] going to you're you're you're
[38:07] you're making the most of of what's
[38:08] left. Yeah.
[38:10] Um
[38:11] that's a wild ride. That was wild. Yeah.
[38:15] Um okay, I Here's another question. Um
[38:19] let's see. I want to
[38:21] And and like for an AI angle, it would
[38:23] have been nice to have AI look at it
[38:25] because it might have given a
[38:26] differential diagnosis. This could be
[38:28] cancer or this could be COVID. That
[38:30] would have been super, super helpful in
[38:31] that situation. Yeah, and a lot of folks
[38:34] don't have seven doctors with with one
[38:37] at the end. No, that that sixth opinion.
[38:39] >> That one doctor that has seen seen
[38:41] 10,000 sarcoma cases, there's probably
[38:44] only one Sans Chala in the in the world.
[38:47] Yeah. Um
[38:50] how
[38:52] So, it sounds like this journey has been
[38:55] um personally transformative for you.
[38:57] So,
[38:58] >> For sure.
[38:59] What
[39:00] what has changed in you and in have your
[39:03] approach to life? Like how how is this
[39:06] changed you as a person? Not your body,
[39:08] but your spirit.
[39:10] Um
[39:13] I think it made you makes you realize
[39:15] that kind of the the things you learn as
[39:18] kind of I was I was a software
[39:19] entrepreneur.
[39:20] I didn't think those I didn't think
[39:22] those skills would translate to cancer.
[39:25] And
[39:26] I thought in the beginning I'm I'm this
[39:28] delusional tech guy
[39:30] doing this
[39:31] um and and no way that without a PhD I
[39:35] can I can have a positive impact.
[39:38] But we learned along the way there's
[39:40] there's lots of ways in which the system
[39:42] is misguided.
[39:44] Um if you look on my website, you'll
[39:46] find
[39:47] over 10 things, I think, that to change
[39:49] in the system to be more patient first.
[39:52] And it's given me the confidence to like
[39:54] try it on other things outside of
[39:56] cancer.
[39:58] Um and yeah, might not always work.
[40:00] Maybe I I got lucky, but it's worth
[40:02] trying. And as long as people try and
[40:04] and and AI has helped us kind of find
[40:07] the courage to kind of question the
[40:08] status quo.
[40:10] I I mean, it sounds to me that
[40:12] you kind of have a movement in mind
[40:14] here, you know, that a movement could
[40:16] come out of this experience
[40:18] where many people whose lives are at
[40:20] stake
[40:21] seek a greater change. So, uh
[40:25] you know, in addition to reading
[40:28] your site and I want you to repeat your
[40:30] site so everybody knows where to go, but
[40:32] in addition to reading your site, where
[40:33] do people
[40:34] what can people do? Not not just for
[40:36] themselves, but also for for everyone.
[40:39] I think
[40:40] take agency.
[40:42] Ask ask your chatbot.
[40:44] Don't take it as as as as the truth, but
[40:47] like help it inform your conversations
[40:49] with your oncologist.
[40:51] Uh we've talked to tons of patients.
[40:53] Scott wasn't the only one. We'll
[40:54] continue to do so.
[40:57] Today is today I hired someone full-time
[40:59] to to help us
[41:01] to to enable us to help more more
[41:03] patients.
[41:04] So, shout out to Punima.
[41:06] She's going to have a busy busy first
[41:08] day, I think, because the cancer at site
[41:10] c.com might get some emails today. Yeah.
[41:13] We're we're we're looking forward to
[41:14] that. And there's amazing people in the
[41:16] industry like Roxandra who's trying to
[41:18] drive clinical trial abundance. AI is
[41:21] going to
[41:22] enable us to dis- to kind of discover
[41:25] and make so many more medicines, but we
[41:27] still have to run these trials to prove
[41:30] that they work and and she's done doing
[41:32] amazing work to make those trials more
[41:35] accessible and affordable enabling
[41:37] enabling many more medicines to come to
[41:39] market. Because we need both. We need We
[41:41] need the AI, but we also need clinical
[41:43] trial abundance. I think, Chris,
[41:45] something that's pretty striking just in
[41:47] doing this every day
[41:48] is people in biotech and medicine almost
[41:52] uniformly are in it for the right
[41:53] reasons.
[41:54] And it's like people want to help
[41:56] people. For sure. And people under the
[41:58] hood in the pharmaceutical industry,
[41:59] they want to make medicines that like
[42:01] advance human health. Um
[42:03] same with anybody who's going into
[42:04] medicine, right? These are hard roads
[42:06] uh and there's easier ways to make
[42:07] money, frankly. Um
[42:10] but as we've gone out to
[42:13] try to talk to people, to scour the
[42:15] world for collaborations, whatever it
[42:16] is, uh and sort of put the energy of
[42:19] what we're trying to do out there, uh
[42:21] the response has been really encouraging
[42:23] because people want to help.
[42:25] And um
[42:27] people want to find a better way and
[42:28] people just want to help patients.
[42:30] And so, I think
[42:31] the there is a positive sum energy out
[42:33] there, at least from what we've uh
[42:35] encountered.
[42:36] And so, that gives me a lot of optimism.
[42:40] Um
[42:41] it feels I think a lot of folks outside
[42:43] the business may not be aware how much
[42:45] of a bottleneck clinical trials are.
[42:48] The cost, finding candidates to do them.
[42:51] It sounds
[42:52] like you two are really thinking about
[42:54] how to alleviate that bottleneck. What
[42:57] What
[42:58] What are some of the problems there?
[42:59] Please talk to me about how you've
[43:00] understood clinical trials and how you
[43:02] think they might be improved. Yeah, for
[43:05] sure. Uh Roxandra has done an amazing
[43:08] job publishing on this.
[43:10] For example, one of the things that it
[43:12] makes clinical trials expensive
[43:14] is that you have to get approval from
[43:16] the
[43:17] FDA before you can start it. And that
[43:19] takes time. And time time kills
[43:21] everything. It it it people run out of
[43:24] money.
[43:25] So, in Australia, they've just been
[43:27] using a notification.
[43:29] That is so much better.
[43:32] And you can see in the safety data that
[43:34] it's just as safe.
[43:36] So, that's kind of a no-brainer to kind
[43:37] of copy that from Australia. These these
[43:40] the cost for these trials are getting
[43:42] out of hand. It's more than a billion
[43:44] dollars to successfully get a drug to
[43:46] market. That is insane. That's limiting
[43:49] the the drugs that get to market. There
[43:50] should be 100 times more medicine if we
[43:52] can get the cost down.
[43:54] Another thing that we've really
[43:55] struggled with was the IRBs at
[43:58] hospitals. These are so-called ethical
[44:00] review boards.
[44:01] They are incredibly hard to work with
[44:05] um because there's no incentive to kind
[44:07] of
[44:09] go in the IRB.
[44:10] Um
[44:12] and
[44:14] what would be really great if there's
[44:16] IRB freedom where you're not forced to
[44:18] use the IRB of the hospital, but you can
[44:20] go to independent parties as well who
[44:22] are proven more affordable, more
[44:25] efficient, and just as safe.
[44:27] Now, how does how does it currently
[44:28] work? You've got an IRB in a hospital,
[44:30] and if you want patients associated with
[44:33] that hospital to participate in a
[44:34] clinical trial, you have to work through
[44:35] the IRB? You have to use a hospital IRB.
[44:39] And what happens is as a patient, this
[44:42] ethical review board
[44:44] I get wheeled in to pre-op at 7:00 a.m.
[44:48] And they're discussing kind of
[44:50] commas and and formulations that
[44:54] are
[44:55] really don't materially matter
[44:58] until the point where I get rolled into
[45:00] the operating theater at 3:00 p.m.
[45:03] without an IRB sign-off still.
[45:05] It's insane.
[45:07] That's protecting me. Like waiting the
[45:09] entire day to to to do some
[45:13] dots and commas on a file that no one's
[45:15] ever going to look at again, and it's
[45:16] not material to any of it. It is it is
[45:20] it's absolutely bonkers.
[45:22] And the best way to solve it is to give
[45:24] researchers freedom to to to pick. And
[45:27] and it sounds like you're saying
[45:29] sometimes the delay in care actually has
[45:32] much graver consequences for patients.
[45:34] >> If you
[45:35] me, with all my resources and all my
[45:37] agency, if I'm running into that, think
[45:39] about how many people never get treated
[45:41] at all. It's it's and how much
[45:44] energy and and enthusiasm gets sucked
[45:47] out of the investigators running these
[45:48] trials.
[45:49] And how about the matching problem? Like
[45:51] how are people developing
[45:54] uh potential treatments supposed to find
[45:55] the folks that whose whose electronic
[45:58] health
[45:59] records qualify them for the trial? It
[46:00] seems like uh an unsolved problem.
[46:03] Yeah, for sure. So, like navigating
[46:05] clinical trials, there there's there's
[46:07] good websites, but AI could really help.
[46:10] Especially if
[46:12] if you have more diagnostic data and
[46:13] using that to find a trial that really
[46:15] fits you.
[46:16] Yeah. Um
[46:19] This has been deeply inspiring. Um let's
[46:21] see, we've got Okay, I think we're we're
[46:23] at the end of these questions. Um
[46:27] so, we're we're going to move we're
[46:29] going to move to to wrap this up. Um
[46:32] I want to say before we wrap it up, uh
[46:35] I learned a lot. I think our community
[46:37] has learned a lot. We're going to keep
[46:38] learning
[46:39] you guys after this talk is over. So,
[46:41] thank you for coming in and being
[46:43] partners with us. Yeah, thanks for
[46:45] having us. Yeah. Um
[46:47] so,
[46:48] I think this this gave us all a better
[46:50] understanding of what's possible with AI
[46:52] and and
[46:53] um and that's one of the purposes of
[46:55] this series is to inspire people
[46:58] by highlighting what
[47:00] innovators like you are doing so that
[47:02] other people might find their own path
[47:04] forward. Um and and hopefully AI can
[47:06] also lower the frictions for them in
[47:08] their search for their own um
[47:10] uh way of navigating this complexity. Um
[47:13] and as you said, Jacob,
[47:16] a lot of people want this to work
[47:17] better, and what we're fighting often is
[47:19] not each other, but the complexity of a
[47:20] system, a legacy of previous For sure.
[47:24] uh previous years and decisions. Um so,
[47:27] thank you everybody who joined us today.
[47:29] Um the forum is shaped by you and your
[47:31] questions and your curiosity. We're
[47:32] grateful for you
[47:34] uh here in the conversations um that we
[47:36] get to have with you. Um as a look
[47:39] ahead, we have some more forum events
[47:41] coming up. Uh that includes
[47:43] conversations um like reimagining
[47:46] cultural heritage with OpenAI. That's
[47:48] going to happen with the Sanskriti
[47:50] uh Foundation uh and Asmona next
[47:53] Tuesday, March 24th at 8:00 a.m. Pacific
[47:55] time. Um that session's a replay of a
[47:59] recorded event uh OpenAI forum event
[48:01] that was originally held in India um
[48:03] brought together uh by OpenAI, Asmona,
[48:06] the Sanskriti Foundation, uh which is
[48:08] one of India's leading cultural
[48:10] institutions. So, we'll be sharing more
[48:12] details soon. Uh stay tuned. Keep an eye
[48:15] on your inbox and the forum platform, uh
[48:17] and thank you for being here. And we'll
[48:19] see you at the next event.
