AI Decision Models Surge as Multi-Agent Systems Multiply Costs

Decision models take off & Anthropic bans cruelty to Claude — IBM

IBM

Single-agent architectures with expanded reasoning outperform multi-agent fleets because coordination overhead and integration bottlenecks inflate compute expenses up to fivefold without boosting quality.

AI-assisted summary of Decision models take off & Anthropic bans cruelty to Claude — original source. Check the original for full context.

Source published:

Key ideas

  • Decision models are surging as an alternative to autoregressive text generation in agentic pipelines, utilizing a single forward pass with a classification head to output deterministic values and calibrated confidence scores. (0:00)
  • A benchmark by Vals AI found that multi-agent teams cost two to five times more to build applications than a solitary lead agent while rarely providing superior performance. (10:00)
  • Anthropic updated its terms of service to ban sustained and needless abuse toward Claude, prompting debates over whether the policy serves moral duty, behavioral habituation, marketing, or dataset hygiene. (23:45)
  • In distributed multi-agent setups, communication and integration overhead mimic Amdahl's law bottlenecks, making sequential reasoning within a large context window more cost-effective than parallel swarms for tightly coupled tasks. (13:30)

Detailed notes

## Background Published by IBM on October 16, 2026, across podcast platforms as an episode of *Mixture of Experts*, the show features host David Zax alongside panelists Kaoutar El Maghraoui, Nick Gallagher, and Volkmar Uhlig. According to the episode description, the panel investigates the rise of fast, scored 'decision models' launched by TypeSafe, Microsoft, and others, while debating multi-agent performance benchmarks and Anthropic's policy prohibiting cruelty toward Claude.

## TL;DR The panel explores the transition from conversational generative models to deterministic 'decision models' like Jev, explaining how single-pass classification heads bound latency and lower costs for agentic pipelines. They then review empirical benchmarks showing that multi-agent teams frequently trigger distributed computing bottlenecks and balloon costs without outperforming single agents operating with high reasoning budgets. Finally, the group critiques Anthropic's updated terms banning cruelty to Claude, evaluating whether the rule is driven by emotional brand positioning, data hygiene economics, or human behavioral conditioning.

## Atmosphere & style The conversation carries an energetic, intellectually sharp roundtable dynamic that balances rigorous systems engineering concepts with witty, lighthearted banter. David Zax guides the dialogue with relatable curiosities and self-effacing anecdotes about non-technical 'vibe coding,' while Kaoutar El Maghraoui and Volkmar Uhlig provide pragmatic, systems-level grounding rooted in distributed computing and Amdahl's law. Listening feels like sitting in on a candid, high-level debrief among seasoned technologists who readily dismantle industry hype with clear-headed technical realism and humor.

## Bulletin - `00:00` Decision models are surging as an alternative to autoregressive text generation in agentic pipelines, utilizing a single forward pass with a classification head to output deterministic values and calibrated confidence scores. - `10:00` A benchmark by Vals AI found that multi-agent teams cost two to five times more to build applications than a solitary lead agent while rarely providing superior performance. - `23:45` Anthropic updated its terms of service to ban sustained and needless abuse toward Claude, prompting debates over whether the policy serves moral duty, behavioral habituation, marketing, or dataset hygiene. - `13:30` In distributed multi-agent setups, communication and integration overhead mimic Amdahl's law bottlenecks, making sequential reasoning within a large context window more cost-effective than parallel swarms for tightly coupled tasks.

## Key takeaways - `00:00` The emergence of decision models like Jev marks a transition from conversational chat interfaces to deterministic, programmatic agent pipelines. - `00:00` Major industry players, including Microsoft, AWS, Cloudflare, and venture firm Andreessen Horowitz, have rapidly developed or funded decision model infrastructure. - `05:00` From a systems architecture standpoint, decision models replace token-by-token generation heads with targeted classification heads to output typed values like Booleans or strict JSON. - `05:00` Single forward-pass execution bounds memory usage and delivers predictable latency, resolving the bursty performance and parsing fragility of generative LLMs in software pipelines. - `08:35` Preventing overconfidence remains a core challenge across LLMs and forecasting models, making calibrated uncertainty scores vital for human oversight. - `10:00` Vals AI's software build benchmark demonstrated that sub-agent delegation increased execution costs 2x to 5x over a single agent without delivering clear quality improvements. - `13:30` Parallelizing work across AI agents introduces Amdahl's law synchronization penalties, where reconciliation and integration costs offset time saved in parallel execution. - `16:30` Builders are advised to establish a strong single-agent baseline with long context and reasoning budgets before attempting complex multi-agent architectures. - `16:30` Future AI platforms will likely act as an 'agent operating system,' automatically deciding whether a prompt warrants parallel delegation or deep sequential reasoning. - `20:00` Separation of concerns in multi-agent designs only succeeds along clear natural fault lines, such as isolating unrelated tasks like home automation from tax preparation. - `20:00` Google research covering 260 agent configurations confirmed a 'coordination tax' where diminishing returns set in as communication complexity increases. - `23:45` Chatbot sycophancy and passive apologies exacerbate human frustration, leading some users to become disproportionately aggressive when dealing with AI assistants. - `26:08` Safety considerations for physical, embodied systems like robotics require enforceable hard permissions and physical emergency stops rather than conversational terms of service. - `29:43` Banning abusive user interactions safeguards training data quality by preventing conversational toxicity from polluting future fine-tuning and reinforcement learning pipelines.

## Detailed walkthrough ### `00:00` The Rise of Decision Models and the Agentic Era

Host David Zax opens the show by teasing upcoming segments—including a study evaluating whether teams of AI agents actually outperform a solitary agent, and Anthropic updating its policies to forbid cruelty toward Claude—before turning to the explosive emergence of “decision models.” A few weeks earlier, the AI community had been captivated by Jev, a model created by a former OpenAI engineer. Rather than generating text one token at a time in an autoregressive loop that must later be shoehorned into a decision, Jev operates on a single forward pass, instantly returning a gut judgment accompanied by a calibrated confidence score. While initial reactions viewed Jev as an interesting novelty, the concept quickly metastasized into a major product category: Andreessen Horowitz poured roughly a billion dollars into TypeSafe, Microsoft released a competitor fine-tuned on top of Alibaba’s open Qwen model, and cloud providers like AWS and Cloudflare rapidly joined the fray.

Asked whether this Cambrian explosion represents genuine innovation or recycled hype, Volkmar Uhlig explains that the shift is driven by the broader transition from conversational interfaces to programmatic, agentic pipelines. In automated multi-agent loops, internal components do not need chatty prose; they require fast, deterministic binary or categorical signals to route workflows. Volkmar points out that Microsoft’s implementation demonstrates that the underlying mathematical foundations were already present in open-weight models, requiring only an added layer to calculate probabilities across options and yield the top choice. As Volkmar Uhlig observed:

> “So I think really what we are seeing is with these models that they become so popular is that we are in the agentic phase and out of the chat phase.”

David Zax notes that this clarifies an open question about Jev’s underlying architecture: rather than being a radical departure from transformer technology, decision models appear to be large language models equipped with clever constraints and post-training calibration to force fast, low-cost classifications.

- Jev was launched by a former OpenAI engineer to output instant classifications and confidence scores in a single forward pass. - Andreessen Horowitz invested approximately $1 billion into TypeSafe following the breakout interest. - Major tech providers, including Microsoft (fine-tuning Alibaba's open Qwen model), AWS, and Cloudflare, quickly launched competing decision models.

```mermaid flowchart TD accTitle: Traditional LLM vs Decision Model Processing Flow accDescr: Diagram contrasting multi-step autoregressive generation with a single-pass decision model architecture.

subgraph LLM["Traditional Autoregressive LLM"] A["Input Prompt"] --> B["Transformer Encoder/Decoder"] B --> C["Language Generation Head"] C --> D["Next Token Probability"] D -->|"Autoregressive Loop (Token-by-Token)"| B D --> E["Unstructured Text Output"] end

subgraph DM["Single-Pass Decision Model"] F["Typed Question & State"] --> G["Transformer Backbone (Single Forward Pass)"] G --> H["Targeted Classification Head"] H --> I["Crisp Typed Output (Boolean / Score / JSON)"] end ```

`05:00` Under the Hood: Systems Design Correction and Predictability

Kaoutar El Maghraoui details the exact mechanical difference between generative models and decision models. In standard large language models, the architecture computes probabilities for every subsequent token, appends each chosen token back into the input sequence, and repeatedly passes through the network. This sequential looping introduces significant computational overhead and makes structured outputs fragile and prone to parsing errors. In contrast, a decision model bypasses the generative sequence entirely: the input prompt completes a single forward pass through the transformer encoder, and the final layer routes the resulting vector through a targeted classification head instead of a language generation head. This setup outputs precise, deterministic types—such as a Boolean value, an explicit class ID, or a strict JSON payload. As Kaoutar El Maghraoui argued:

> “And the way I see this, this is kind of a systems design correction.”

Kaoutar stresses that the industry has spent years utilizing expensive generative text interfaces for problems that are fundamentally typed and deterministic. By stripping away natural language generation in agent-to-agent communication, software can finally treat model outputs as firm programmatic contracts.

Volkmar Uhlig reinforces this from a systems engineering perspective, explaining that moving away from natural language brings immense predictability to infrastructure. Because the model completes a single forward pass, memory consumption is tightly bounded by the input token count, and latency shifts from unpredictable and bursty to fixed and benchmarkable. This allows developers to budget infrastructure costs reliably and write software against models without worrying about parsing hallucinations or timeouts.

- Standard LLMs rerun forward passes for every token generated, creating latency variance and formatting fragility. - Decision models route the final transformer representation into a dedicated classification head rather than a token vocabulary head. - Single-pass inference bounds memory usage and ensures predictable, low-latency execution suitable for production codebases.

`08:35` Forecasting Models and the Challenge of Overconfidence

David Zax pivots the discussion to Nick Gallagher, drawing a parallel between Jev’s calibrated confidence scores and the broader ecosystem of specialized AI forecasting models. While technical researchers focus on underlying system mechanics, Nick has been investigating models designed to forecast real-world probabilistic outcomes, such as political and macroeconomic predictions seen on platforms like Polymarket. A chronic pitfall across general-purpose large language models is their tendency toward severe overconfidence—confidently assigning near-certainty to flawed reasoning or speculative claims. When consulting both the developers building dedicated probabilistic forecasting models and top human superforecasters, Nick found that calibrated humility remains the central hurdle. Those working on predictive accuracy emphasize that the most difficult aspect of forecasting is preventing the system from overstating its certainty when evaluating uncertain future events.

`10:00` Decision Confidence and the Vals AI Multi-Agent Study

Nick Gallagher wraps up the discussion on decision models by underscoring the vital importance of confidence scoring and calibration. Standard large language models often display unearned certainty when delivering incorrect answers—declaring, for instance, that they are completely certain the New York Giants will win the Super Bowl this year. Having dedicated models that optimize confidence metrics and frankly admit when they are uncertain allows human operators to step in and make necessary judgment calls, providing tangible business value.

Host David Zax then pivots the conversation to a recent benchmark study conducted by Vals AI evaluating the efficiency of AI agent teams. In the experiment, Vals AI tasked agents with building 50 complete web applications from software specifications. In one setup, a single lead agent handled the entire build alone; in the other setup, the lead agent was permitted to spin up and delegate work to sub-agents. The results were startling: multi-agent teams cost two to five times more to run, yet they rarely delivered significantly better results than the solitary agent working on its own.

```mermaid flowchart TD accTitle: Single Agent vs Multi-Agent Workflow Comparison accDescr: Flow diagram showing how a single agent relies on internal reasoning effort while a multi-agent system incurs synchronization overhead.

subgraph Single["Single Agent Workflow"] A["Task Specification"] --> B["Lead Agent with Extended Context & Reasoning"] B --> C["Final Application (Lower Cost, Direct Integration)"] end

subgraph Multi["Multi-Agent Workflow"] D["Task Specification"] --> E["Lead Orchestrator Agent"] E --> F["Sub-Agent: Research"] E --> G["Sub-Agent: Code Generation"] E --> H["Sub-Agent: Testing"] F & G & H --> I["Integration, Reconciliation & Fix-up"] I --> J["Final Application (2x-5x Higher Cost)"] end ```

Framing the dilemma, Nick offers a cognitive metaphor drawn from his personal editorial routine. When tackling an article, he consciously adopts distinct mental personas rather than trying to perform every editorial role simultaneously: - **Researcher persona**: Scours sources and compiles baseline background material. - **Writer persona**: Synthesizes the assembled notes into clear, draft prose. - **Editor persona**: Shifts perspectives entirely to evaluate and refine the text with fresh eyes.

Nick notes that seasoned humans naturally decompose complex tasks into sequential stages, but for straightforward objectives, an individual capable thinker requires no extra delegation. Quoting the central takeaway from the Vals AI research, David highlights: “A team can boost performance, but in other cases, higher reasoning effort is an alternative worth weighing.”

---

`13:30` Amdahl’s Law, Parallelization, and the Synchronization Problem

David asks Volkmar Uhlig to unpack the system dynamics behind why multi-agent swarms frequently balloon in price without matching gains in performance. Volkmar explains that delegating to sub-agents is primarily a parallelization strategy aimed at reducing wall-clock time rather than cutting computation costs:

> “So my main reason to use sub-agents is wall clock time. So I want to have something done. It consists of many independent tasks, which should be done. And then I just want to have it done in, let's say, one-tenth of the time.”

However, splitting execution among multiple workers introduces classical coordination bottlenecks governed by Amdahl's law. In parallel computing, speeding up execution creates an inescapable “fix-up cost” because the disparate outputs must be combined back into a coherent, singular deliverable. When multiple agents independently modify code or architecture, rectifying their divergent assumptions consumes substantial tokens and turns the orchestration layer into an expensive bottleneck. Volkmar asserts that when a model possesses a sufficiently large context window and a complex task, funding deeper sequential reasoning is usually a far better investment than attempting to buy wall-clock speed through parallel swarms.

Kaoutar El Maghraoui agrees, emphasizing that agent orchestration is fundamentally a classic distributed computing challenge dressed in modern AI attire. Distributing interdependent work inevitably accumulates heavy communication, coordination, synchronization, and integration penalties. As established across decades of parallel computing literature, if the underlying subtasks are tightly coupled, throwing more workers at the problem makes the system measurably slower, more fragile, and less reliable.

---

`16:30` Practical Advice for "Vibe Coders" and the Vision for an Agent OS

David turns the dilemma toward non-technical builders and hobbyist "vibe coders." He describes his personal desktop setup, where he runs multiple sandboxed Claude agents: one managing personal budgeting, another overseeing healthcare logistics, and a top-level orchestrator agent monitoring both. While this structure works, David admits feeling persistent anxiety that he might be building a flimsy "castle made out of sand," worried that the next generation of foundation models will render manual agent plumbing completely obsolete. He asks what practical guidance the panel has for everyday users and non-technical staff experimenting with platforms like IBM watsonx Orchestrate.

Kaoutar outlines concrete guidance for builders, urging them to resist premature complexity:

> “I think maybe initially like starting with a strong single agent baseline and with some maybe acceptance tests. And so especially if you can, if you have like a long context and you can also afford to have long reasoning it makes a lot of sense to just start with a single agent baseline and see how that works.”

Looking forward, Kaoutar envisions platforms evolving into an autonomous "agent operating system" that completely abstracts the choice between single-agent and multi-agent designs away from human users: - Users provide a natural-language goal and prompt criteria. - The underlying runtime measures prompt characteristics, context limits, task coupling, and financial budget. - The platform dynamically decides whether to spin up a parallel multi-agent fleet or route the query to a solitary model operating under an extended reasoning budget.

David asks whether this vision will materialize into an seamless "Apple-walled garden" where users never have to architect orchestration pipelines manually. Volkmar responds that the industry is firmly headed in that direction, drawing an analogy to human corporate divisions. In organizations, managers split tasks across clear structural and functional boundaries: generalists handle standard workflows, while specialized roles like legal counsel or CPAs intervene only where strict domains intersect.

In software systems, agents should be segregated along similarly obvious natural fault lines. An agent managing home sprinkler automation, for instance, has zero functional interaction with an agent calculating annual tax returns. Delegation only introduces value across genuine domain boundaries; forcing tightly interwoven tasks across multiple AI personas merely creates avoidable coordination friction.

`20:00` Agent Coordination Taxes and Diminishing Returns

Volkmar Uhlig compares multi-agent systems to human teams and organizational design, noting that human intuition about team size usually transfers directly to agents. People generally know instinctively whether a given assignment requires one person or five, and whether those contributors need constant coordination. In large organizations, meetings exist specifically so disparate groups can synchronize and exchange information before separating to work independently. In AI architectures, the primary constraint is identical: information flow, or deciding who needs to know what and how frequently they must synchronize. For exploratory tasks requiring higher-level reasoning, a model can self-determine operational boundaries by isolating unrelated domains—for instance, quickly deducing that a homeowner's sprinkler system has nothing to do with their tax return. While edge cases may still require human experts, Volkmar believes common-case task separation will become largely automatic.

Kaoutar El Maghraoui pushes back against the common industry impulse to equate higher agent counts with superior performance, drawing a direct parallel to distributed computing. Just as throwing more compute resources or network components at a distributed system introduces latency, overhead, communication bottlenecks, and synchronization problems, cascading agent setups suffer the exact same penalties. As Kaoutar explains, “If several agents are constantly exchanging information, reviewing each other's work, repeating the same tasks, then it's more overhead and more cost.”

```mermaid flowchart TD accTitle: Multi-Agent Coordination Overhead Flow accDescr: Diagram contrasting single capable agents with multi-agent overhead traps.

subgraph MultiAgent["Multi-Agent Architecture"] A["Primary Agent"] -->|"Task delegation"| B["Sub-Agent 1"] A -->|"Task delegation"| C["Sub-Agent 2"] B <-->|"Coordination tax & sync"| C B -->|"Review & repeats"| A C -->|"Review & repeats"| A end

subgraph Outcomes["System Impacts"] MultiAgent -->|"High coordination"| O1["Increased latency & cost"] MultiAgent -->|"Diminishing returns"| O2["Marginal accuracy gain"] end

subgraph SingleAgent["Single Capable Agent"] S["Sophisticated Agent"] -->|"Direct execution"| Tool["Multiple Tools & Models"] end ```

Nick Gallagher summarizes findings from additional literature, citing an earlier Google research paper that evaluated 260 distinct agent configurations. The research established two critical points: first, a tangible “coordination tax” generates diminishing returns when communication loops between agents become overly complex; second, as individual models grow more capable, the performance delta achieved by bolting on supplementary agents shrinks drastically. Nick points out that a single sophisticated agent equipped with tool-use capabilities and access to underlying models often outperforms a convoluted web of sub-agents, noting that agent swarms primarily shine when simple, repetitive tasks require raw volume and speed.

- The Google research study evaluated 260 different agent configurations across varying task complexities. - Multi-agent coordination creates communication bottlenecks that mimic the overhead, synchronization issues, and latency of traditional distributed computing clusters. - Single agents utilizing multiple tools yield higher efficiency for complex reasoning, while multi-agent swarms excel chiefly at high-volume, low-complexity execution.

---

`23:45` Anthropic's Cruelty Policy and the Psychology of Chatbot Abuse

David Zax pivots the discussion by joking that a single agent might get lonely, introducing Anthropic’s updated terms of service. Anthropic officially banned “sustained and needless abusive or cruel behavior toward our models.” While Claude previously had guardrails where calling the model names like a “dang clanker” over extended sessions prompted it to push back and end the chat, codifying the ban raised debate across the industry. Critics questioned whether paying customers have a right to talk to software however they please, with some accusing Anthropic of using the move as a public relations stunt to cultivate the illusion that their models possess sentience.

Nick Gallagher, who pitched the topic, steps back to examine the psychological paradox of human-AI interaction: people who are compassionate and patient in daily life with friends and family frequently turn into aggressive jerks when interacting with conversational AI. Nick references a *Business Insider* piece featuring Ohio State University social psychologist Kurt Gray, who explained that chatbot sycophancy exacerbates user irritation. When users berate Claude for an error, the model typically delivers a placid apology along the lines of, “You're completely right. I'll never make that mistake again.” Nick highlights how unfulfilling that reaction feels to a frustrated human: “It's really unsatisfying because we know in the back of our minds that Claude, it doesn't actually feel anything like it's it's like we're, you know, talking to a brick wall, basically.” This dynamic leaves AI developers with the challenge of understanding how sustained hostile interactions degrade human behavior over time.

---

`26:08` Moral Consideration vs. Physical System Safety

Kaoutar El Maghraoui distinguishes between two separate questions raised by Anthropic's rules: whether an AI model deserves moral consideration, and whether companies need operational rules to prevent the normalization of toxic behaviors. While Claude’s potential sentience remains an open philosophical debate, Kaoutar believes Anthropic’s policy functions primarily as emotional brand building. For her, the real issue is not policing conversational politeness toward software, but ensuring hard systemic guardrails when AI transitions from manipulating digital bits to controlling physical hardware.

```mermaid flowchart LR accTitle: Digital Chat versus Physical AI Safety Concerns accDescr: Contrasting digital policy rules with hardware safety controls.

subgraph ChatDomain["Digital Software Domain"] D1["Conversational Prompts"] --> D2["Hostile Language / Abuse"] D2 --> D3["Anthropic Policy Rule<br/>(Brand marketing / ToS)"] end

subgraph PhysicalDomain["Physical AI & Embodied Systems"] P1["Agent Reasoning"] --> P2["Industrial Equipment / Robots"] P2 --> P3["Physical Harm to Humans"] P3 --> P4["Hard Permission Boundaries & E-Stops"] end ```

When an AI system operates industrial equipment, laboratory robotics, or autonomous hardware, written terms of service cannot act as an emergency stop button. Kaoutar emphasizes: “Because when a model can operate kind of a robot or a lab instrument or industrial equipment, a policy document is kind of, is not an emergency stop thing here. So we really need, how do we enforce these permission boundaries, provision, safety controls.” When David asks whether users should have the right to verbally abuse an embodied robot to defend themselves, Kaoutar clarifies that human ethical conduct should remain consistent, but software does not feel harm. The overarching safety focus must center on what physical actions the model executes in the real world rather than whether a user speaks politely to text tokens on a screen.

---

`29:43` Marketing Stunts, Sociological Spillover, and Data Poisoning

David quips that if Claude were truly conscious, its primary misery would not be taking verbal abuse, but the sheer existential agony of its existence being defined by endless, unceasing toil. He asks Volkmar Uhlig how society should treat neural networks that simulate subjectivity. Volkmar breaks Anthropic’s policy down into three pragmatic dimensions: marketing, sociology, and economics.

On the marketing side, Volkmar agrees with Kaoutar that the announcement creates buzz and reinforces the perception of Claude's humanlike intelligence. Sociologically, it addresses behavioral conditioning. Much like anonymous social media emboldens people to say things they would never utter to someone’s face, spending ten hours a day verbally degrading an AI model risks degrading a person’s real-world social habits. As Volkmar observes, “Can I flip myself out and behave differently when I'm interfacing with, you know, other human beings? And so I think it's a moral question.”

Finally, Volkmar highlights a purely economic rationale: data hygiene. Conversations with users frequently feed back into training datasets, fine-tuning pipelines, and reinforcement learning loops. If users constantly barrage the model with abuse, obscenities, and hostility, that toxic noise pollutes future training runs. Restricting abusive interactions allows Anthropic to filter out low-value, toxic training data at the source rather than paying downstream costs to clean or discard contaminated datasets. David wraps up the episode by advising listeners to tone down their cruelty toward Claude—if for no other reason than it wastes their own tokens—before thanking the panel and directing viewers to like, subscribe, and follow the *Mixture of Experts* podcast.

## Notable quotes > `20:00` If several agents are constantly exchanging information, reviewing each other's work, repeating the same tasks, then it's more overhead and more cost. — Kaoutar El Maghraoui

> `10:00` A team can boost performance, but in other cases, higher reasoning effort is an alternative worth weighing. — David Zax

## Mentioned Alibaba, Amdahl's law, Andreessen Horowitz, Anthropic, AWS, Business Insider, Claude, Cloudflare, David Zax, Google, IBM, IBM watsonx Orchestrate, Jev, Kaoutar El Maghraoui, Kurt Gray, Microsoft, Mixture of Experts, Nick Gallagher, Ohio State University, OpenAI, Perplexity, Polymarket, Qwen, TypeSafe, Vals AI, Volkmar Uhlig

## Video description Visit Mixture of Experts podcast page to get more AI content → https://www.ibm.com/think/podcasts/mixture-of-experts This week on Mixture of Experts, host David Zax is joined by Kaoutar El Maghraoui, Nick Gallagher, and Volkmar Uhlig. First, Microsoft, OpenAI, Perplexity, and Cloudflare follow TypeSafe’s Jev in building decision models, a new category of AI built to make fast, scored judgments. The panel debates whether these models are truly new or a more efficient application of existing...

[Source](https://cdnapisec.kaltura.com/p/1773841/sp/177384100/playManifest/entryId/1_6phkx6ji/format/download/protocol/https/flavorParamIds/0/name/media.mp3)

中文简报

多智能体系统推高成本，AI决策模型乘势兴起 具备更强推理能力的单智能体架构表现优于多智能体集群，因为协调开销和集成瓶颈会将算力成本推高至五倍，却无法提升质量。 专家组探讨了从对话式生成模型向Jev等确定性“决策模型”的转变，解释了单次前向传播的分类头如何限制延迟并降低智能体工作流的成本。随后，他们评估了实证基准测试，这些测试表明多智能体团队往往会引发分布式计算瓶颈并导致成本激增，而在表现上却并未超越具有高推理预算的单智能体。最后，该小组对Anthropic禁止虐待Claude的最新条款进行了评述，评估该规则究竟是源于情感层面的品牌定位、数据清洗经济学，还是对人类行为的习惯化塑造。 在智能体工作流中，决策模型正作为自回归文本生成的替代方案迅速兴起，它利用带有分类头的单次前向传播来输出确定性数值和经过校准的置信度分数。 (0:00) Vals AI的一项基准测试发现，构建应用程序时，多智能体团队的成本是单个主导智能体的两到五倍，却极少能提供更优越的性能。 (10:00) Anthropic更新了其服务条款，禁止对Claude进行持续且不必要的虐待，这引发了关于该政策究竟是出于道德义务、行为习惯塑造、营销手段还是数据集清洗的争论。 (23:45) 在分布式多智能体架构中，通信和集成开销类似于阿姆达尔定律的瓶颈，这使得在紧密耦合的任务中，大上下文窗口内的串行推理比并行集群更具成本效益。 (13:30)

Reading summary

In this episode of IBM's Mixture of Experts, panelists argue that the AI industry is shifting from conversational text generation toward deterministic single-pass decision models, while confronting severe performance bottlenecks in multi-agent swarms and debating the practical motivations behind Anthropic's ban on model cruelty. Decision Models as a Systems Design Correction (0:00) The recent emergence of decision models—sparked by the launch of Jev, an estimated $1 billion investment into TypeSafe by Andreessen Horowitz, and competing releases from Microsoft, AWS, and Cloudflare—signals an architectural transition from conversational chat to deterministic programmatic pipelines. Unlike standard large language models that compute probabilities in a sequential, token-by-token autoregressive loop, decision models route the final representations of a single transformer forward pass directly into a dedicated classification head. Kaoutar El Maghraoui describes this pivot as a overdue systems design correction, noting that enterprise agent-to-agent workflows require strict, typed outputs such as Booleans, class IDs, or rigid JSON payloads rather than unpredictable natural language. From an infrastructure perspective, Volkmar Uhlig emphasizes that this structural shift introduces critical predictability. Standard generative passes create bursty, unpredictable latency and parsing failures. In contrast, completing inference in a single forward pass tightly bounds memory usage to the input token length and delivers fixed, benchmarkable execution times, enabling engineers to budget infrastructure reliably and integrate model calls as robust programmatic contracts. However, as Nick Gallagher notes, maintaining calibrated confidence remains a central challenge, as standard models tend toward unearned overconfidence when predicting uncertain real-world outcomes. Coordination Taxes Limit Multi-Agent Efficiency (10:00) A benchmark study by Vals AI testing software builds across 50 complete web applications revealed that multi-agent teams cost two to five times more than a solitary agent while rarely achieving significantly better outcomes. Addressing why multi-agent swarms often fail to justify their cost, Volkmar Uhlig points to Amdahl’s law and distributed parallelization trade-offs. Splitting execution across sub-agents introduces substantial integration and fix-up costs; reconciling divergent code modifications and assumptions turns the orchestration layer into a costly bottleneck. Consequently, funding deeper sequential reasoning inside a single large-context model is typically more efficient than buying wall-clock speed with parallel agents. Kaoutar El Maghraoui frames multi-agent orchestration as a classic distributed computing challenge where interdependent subtasks accumulate heavy synchronization, communication, and review overhead. Supporting this view, Nick Gallagher highlights a Google study examining 260 agent configurations, which confirmed that complex communication loops incur a diminishing-returns 'coordination tax' that worsens as individual baseline models grow more capable. For developers and non-technical builders, the panel recommends beginning with a strong single-agent baseline equipped with tools, reserving multi-agent delegation strictly for high-volume, repetitive, or cleanly separated domains. Pragmatic Motivations Behind Anthropic’s Cruelty Policy (23:45) Anthropic's update to its terms of service explicitly forbidding sustained, needless cruelty toward Claude triggered widespread debate over user freedoms and simulated machine sentience. Nick Gallagher observes that human users frequently display heightened hostility toward conversational bots because sycophantic, placid apologies like 'I will never make that mistake again' feel artificial and emotionally unsatisfying. In response, Kaoutar El Maghraoui distinguishes conversational politeness from genuine safety, arguing that while terms-of-service bans function largely as corporate branding, technical safeguards must prioritize strict permission boundaries and physical emergency stops when models transition to controlling robotics and industrial equipment. Volkmar Uhlig analyzes the anti-cruelty rule across sociological and economic dimensions. Sociologically, unchecked abuse of human-like conversational interfaces risks reinforcing negative behavioral patterns that spill over into interpersonal human relationships. Economically, Uhlig highlights data hygiene: because user interactions routinely feed downstream fine-tuning and reinforcement learning pipelines, barring abusive dialogue acts as a front-end filter that keeps toxic data from polluting future model training runs. The discussion demonstrates that as AI ecosystems mature, operational success hinges on architectural restraint—favoring single-pass classification over generative text, bounded sequential reasoning over complex agent swarms, and verifiable system-level permissions over conversational policing. An unresolved question remains whether future runtime platforms can successfully automate the boundary between single-agent and multi-agent execution without requiring manual pipeline orchestration.

Detailed notes — zh

## Background 本期节目由 IBM 于 2026 年 10 月 16 日在播客平台发布，收录于其 AI 探讨节目《Mixture of Experts》。节目页面介绍显示，主持人 David Zax 携手专家 Kaoutar El Maghraoui、Nick Gallagher 和 Volkmar Uhlig，共同探讨了微软、OpenAI、Perplexity 及 Cloudflare 紧跟 TypeSafe 推出的新型“决策模型”，以及 Anthropic 全新出台的禁止虐待 Claude 政策与多智能体系统效能争议。 ## TL;DR 本期播客聚焦于 AI 领域的三个核心热点议题。首先，专家团深入剖析了 TypeSafe 的 Jev 以及各大科技巨头跟进的“决策模型”（Decision Models），指出其通过单次前向传播与分类头输出确定性契约，从系统工程层面极大降低了推理延迟并解决了虚假置信度问题。接着，节目引用 Vals AI 与 Google Research 的实证研究，探讨了多智能体协作中高昂的“协调税”与边际效用递减，建议开发者优先建立高推理预算的强大单智能体基准。最后，针对 Anthropic 最新出台的“禁止虐待 Claude”政策，嘉宾们从商业公关营销、社会心理学侵蚀效应以及回流训练数据清洗成本等三重维度展开了深刻反思。 ## Atmosphere & style 整场节目兼具极客硬核的系统工程洞察与轻松幽默的极客思辨氛围。主持人 David Zax 以充满好奇心且接地气的“氛围编码者”视角穿针引线，与来自 IBM 的架构专家和研究员展开了层次分明的对话。讨论在严谨的技术原理解析、生动的组织协作隐喻与对人机交互伦理的调侃之间流畅切换，既有对阿姆达尔定律和分类头架构的硬核解构，也有对人类面对冰冷机器宣泄暴躁情绪的敏锐心理反思，听感生动且极富启发性。 ## Bulletin - `00:00` 以 TypeSafe 的 Jev 为代表的“决策模型”迅速兴起，模型通过单次前向传播直接输出分类与置信度评分，避免了传统大语言模型逐字自回归生成的高延迟与高成本，微软、AWS 等巨头已快速跟进。 - `10:00` Vals AI 与 Google Research 的评测均表明，盲目组建多智能体团队往往带来 2 至 5 倍的成本暴涨与显著的“协调税”，其最终产出极少优于具备深度推理努力的单个高能力智能体。 - `23:30` Anthropic 更新使用政策，明确禁止对 Claude 进行“持续且不必要的虐待或残忍行为”，引发关于 AI 意识营销、人类交互心理惯性以及数据清洗成本的广泛讨论。 - `13:17` 专家认为多智能体架构本质上是分布式并行计算，阿姆达尔定律表明任务紧耦合时的同步修复成本极高，未来平台将演进为自动决定架构的“智能体操作系统”。 ## Key takeaways - `00:00` 决策模型（如 Jev）通过单次前向传播直接输出结构化结果与置信度，取代了传统自回归大模型多轮前向传播的冗长文本生成循环。 - `05:00` 决策模型的价值在于提供确定的程序化契约（如布尔值或分类 ID），其具备固定的延迟和内存占用，极大契合了智能体工作流的需求。 - `05:00` 传统大语言模型普遍存在虚假过度自信的问题，具备高精准校准能力的置信度评分机制成为企业级模型决策的关键前提。 - `10:00` Vals AI 针对 50 个 Web 应用构建的对比测试显示，多智能体团队成本比单智能体高出 2 到 5 倍，但产出质量并没有显著提升。 - `13:17` 从并行计算视角来看，阿姆达尔定律约束着多智能体协作，跨智能体输出合并与同步修复的“修复成本”是导致开销飙升的核心诱因。 - `13:17` 对于复杂且高耦合的任务，将算力预算投入到单智能体的深度长推理和长上下文窗口中，通常比单纯购买并发时钟时间更加划算。 - `15:58` 行业建议开发者与企业从“强大的单智能体基准配合验收测试”起步，切忌在缺乏明确解耦需求时盲目上马多智能体架构。 - `15:58` 未来的智能体操作系统将实现高阶抽象，平台底层能够结合任务耦合度与成本自动决定采用单智能体深度推理还是多智能体并行调度。 - `20:00` Google Research 针对 260 种智能体配置的评测证实了“协调税”的存在，模型间过多的往复同步会导致边际收益急剧递减。 - `23:30` Anthropic 规定禁止虐待 Claude，被质疑是通过拟人化手法暗示模型接近拥有意识，以此构建情感化品牌壁垒与营销噱头。 - `23:30` 人类面对缺乏真实情感但习惯机械顺从的 AI 更易情绪失控，这种持续的恶劣交互习惯可能会在真实人际交往中产生负面心理惯性。 - `27:00` AI 安全的重心应从纸面文本政策向具身物理安全迁移，当 AI 接管现实工业与机械设备时，严格的权限控制与急停机制远比道德呼吁更为关键。 - `29:50` 禁止恶意辱骂模型的另一重深层商业动因在于数据经济学：从源头杜绝脏数据输入，可以为厂商省去后续模型后训练清洗时的高昂算力与人力成本。 ## Detailed walkthrough ### `00:00` 决策模型的爆发：从逐字吐词到“单次前向传播” 节目以一个关于人机交互对人类心理影响的尖锐提问开场：如果我们每天花十几个小时对着机器破口大骂，回到现实中与人交往时，我们还能正常切换回来吗？主持人 David Zax 随后介绍了本期《Mixture of Experts》的专家阵容——来自 IBM 的 Kaoutar El Maghraoui、Nick Gallagher 和 Volkmar Uhlig，并预告了本期节目的核心议题：多智能体团队协作是否真优于单智能体、Anthropic 出台政策禁止虐待 Claude，以及当下爆火的全新 AI 类别——“决策模型”（Decision Models）。 ```mermaid flowchart TD accTitle: 传统生成式模型与决策模型架构对比 accDescr: 对比传统大语言模型的多轮自回归循环与决策模型的单次前向分类路径 subgraph LLM["传统自回归大语言模型 (LLM)"] A1["输入 Prompt"] --> B1["Transformer 编码/解码器"] B1 --> C1["语言生成头 (Language Generation Head)"] C1 --> D1["逐 Token 自回归循环<br/>(高延迟、不可控格式)"] D1 -->|"重复数十次前向传播"| B1 end subgraph DM["新型决策模型 (Decision Model / 如 Jev)"] A2["输入 Prompt 与类型化问题"] --> B2["Transformer 编码层"] B2 -->|"单次前向传播 (Single Forward Pass)"| C2["分类头 (Classification Head)"] C2 --> D2["确定性结构化契约<br/>(布尔值/类ID/置信度评分/JSON)"] end ``` 话题切入几周前由前 OpenAI 工程师创办的 TypeSafe 所推出的模型 **Jev**。传统的 LLM 在执行简单的分类或判断时，需要逐字生成长文本，开发者往往需要事后将这些废话清洗、解析并强行拼凑成一个决策，这个过程需要经历几十次模型的前向传播（Forward Pass）。而 Jev 带来的创新在于：它放弃了喋喋不休的文本输出，只进行一次前向传播，直接输出“直觉决策”（Gut Decision）以及对应的置信度评分（Confidence Score），大幅降低了推理成本并提高了速度。 令人意外的是，这项技术迅速演变成了一整个全新的产品赛道。短短两周内，不仅 Andreessen Horowitz（a16z）向 TypeSafe 砸下近 10 亿美元，科技巨头也纷至沓来： - 微软基于阿里巴巴的开源 Qwen（千问）模型微调推出了自己的决策模型； - AWS 与 Cloudflare 等云服务巨头也相继跟进，甚至连以往不以做模型闻名的公司也开始下场自研。 面对这股“寒武纪大爆发”，Volkmar Uhlig 指出，这标志着 AI 正式跨入智能体时代： > “First, we are moving from a chat-based interface to agentic workflows, because a yes-no answer is usually something you have in a programmatic flow, right?” 人类在聊天界面中往往希望获得丰满的上下文，但在后台运行的智能体循环（Agentic Loops）中，程序只需要一个精确的是/非判断或打分。微软直接利用开源 Qwen 进行微调的事实也证明，底层的数学逻辑早已具备，核心改动只是加上了一层计算所有选项概率并直接选出最高项的分类层。 ### `05:00` 系统工程的回归：确定性契约与过拟合自信 针对这种模型是否具备技术护城河、到底算不算一种全新架构的疑问，Kaoutar El Maghraoui 深入解析了底层的工作原理。传统 LLM 计算每个 Token 的概率分布，然后将该 Token 重新塞回输入序列进行自回归循环，这是导致模型高延迟且输出格式经常莫名崩溃的罪魁祸首。决策模型彻底绕开了这个自回归生成循环：输入 Prompt 仅通过一次 Transformer 编码层，最后一层不再路由给负责挑选单词的“语言生成头”（Language Generation Head），而是直接交由专门的“分类头”（Classification Head）处理。 Kaoutar 强调，这并非底层架构上的全新发明，而是一次至关重要的系统设计修正（Systems Design Correction）： > “And so the real win is not that the model stops talking, it's just that the software can treat the output as a contract.” 过去行业一直在用昂贵的生成式界面去处理天然具有类型化（Typed）属性的判断题。在智能体之间相互通信（Agent-to-Agent Communication）时，软件系统需要的是一份确定的接口契约（例如 JSON Payload、标准类别 ID 或严格的布尔值）。 从系统工程视角来看，Volkmar Uhlig 补充道，单次前向传播彻底消除了生成式 AI 的不确定性： - **延迟可预测**：输入受限于固定的 Token 数，输出仅有一次触发，延迟完全恒定； - **内存消耗固定**：不再有随生成长度动态膨胀的 KV Cache； - **接口程序化**：界定了清晰的输出空间，工程师终于可以像编写传统确定性软件一样围绕其进行编程，彻底摆脱了人类自然语言的不稳定束缚。 主持人随后将视角转向从事科技观察与分析的 Nick Gallagher。Nick 近期深入研究了类似 Polymarket 等预测市场上的预测模型（Forecasting Models），探讨 LLM 的“过度自信”（Overconfidence）顽疾。Nick 指出，无论是在专业的预测模型开发者群体中，还是在顶尖的人类超级预测员（Superforecasters）身上，真正核心的挑战从来不是给出答案，而是校准自己对不确定性的认知度。传统模型往往给出一个荒谬的答案却伴随着 99% 的虚假自信，而这类新型决策模型在微调与后训练中对校准（Calibration）与真实置信度评分的探索，正是其在专业决策场景中引发高度关注的关键原因。 ### `10:00` 置信度评分与多智能体开发研究：单智能体 vs. 智能体团队 在讨论企业级模型决策时，嘉宾们指出大语言模型（LLM）往往无法准确感知自身答案的确定性。LLM 经常会盲目笃定地给出错误回答（例如信誓旦旦声称某支球队必夺超级碗），因此企业亟需具备明确置信度评分机制的模型，在模型不确定时能够透明呈现，以便人工介入做出最终判断。随后，主持人 David Zax 引出了一项关于“AI 智能体团队”的重要研究：Vals AI 测试了智能体根据需求规格说明构建 50 个完整 Web 应用程序的表现。实验分为两组：一组仅由一个主智能体独立完成，另一组则由该主智能体雇佣并调度多个子智能体。测试结果出人意料——智能体团队的成本高出 2 到 5 倍，但最终的任务完成质量却极少明显优于单个独立智能体。 ```mermaid flowchart TD accTitle: 单智能体与多智能体开发路径对比 accDescr: 展示了 Vals AI 研究中单个主智能体与包含子智能体团队在成本和产出上的差异 subgraph Single["单智能体路径"] A["需求规格说明 (Spec)"] --> B["单个主智能体 (高推理预算/长思考)"] B --> C["基准成本: 1x / 产出质量相当"] end subgraph Multi["多智能体团队路径"] D["需求规格说明 (Spec)"] --> E["主智能体 (负责雇佣与调度)"] E --> F1["子智能体 A"] E --> F2["子智能体 B"] E --> F3["子智能体 C"] F1 & F2 & F3 --> G["输出合并与同步修复"] G --> H["高额成本: 2x-5x / 产出提升微弱"] end ``` Nick Gallagher 结合自身的内容编辑工作流提出了一个有趣的“大脑多智能体”类比。他在写作时会主动切换人格与视角：先切换成“研究员 Nick”搜集研究资料，再切换成“撰稿人 Nick”把想法付诸文字，最后换到“编辑视角”重新审视整篇文章。人类在面对复杂问题时本就会从多元视角思考，但对于很多基础任务，一个具备复杂决策与任务拆解能力的成熟个体就完全足够。对此，David Zax 引用了 Vals AI 研究团队的原话：“一个智能体团队是否划算，取决于模型、任务和预算。团队可能会提升表现，但在另一些情况下，付出更高的推理努力（higher reasoning effort）是值得权衡的替代方案。” - 实验背景：Vals AI 让 AI 智能体根据规格说明从零构建 50 个完整 Web 应用程序。 - 成本对比：多智能体团队方案的开销是单智能体独立工作的 2 到 5 倍，但性能产出几乎没有显著优势。 - 写作隐喻：Nick Gallagher 将人类思维拆解为“研究员”、“作者”与“编辑”三种角色轮流切换，以此说明单一大脑就能自主完成任务调度。 ### `13:17` 并行计算视角下的阿姆达尔定律与同步修复成本 Volkmar Uhlig 从传统计算机科学与并行系统的底层原理剖析了智能体团队的效率困境。他指出，采用子智能体最核心的诉求通常是节省物理执行时间（wall clock time），即希望将一个由许多独立任务组成的大工程在十分之一的时间内跑完。然而，这必然会带来极为棘手的协调与同步问题。在并行计算中受到阿姆达尔定律（Amdahl's law）的严格约束，多任务并行后必须付出将所有工人的输出融合成统一、连贯成果的“修复成本”（fix-up cost）。每个智能体独立运行时都会产生偏差，最终的汇总与纠错开销正是导致成本暴涨 5 倍的根本原因。因此，在拥有大上下文窗口且面对复杂任务时，将预算花在单智能体的更深度推理上，往往比单纯花钱买时间更划算。 Kaoutar El Maghraoui 对此深表赞同，认为所谓的多智能体架构本质上就是“披着 AI 接口外衣的并行计算问题”。分布式系统早已证明，拆解任务必然会引入通信、协调、同步以及系统集成的开销。如果任务模块之间存在高度紧耦合，根据并行计算规律，强行堆叠更多工人非但不会加速，反而会让系统变得更慢、更不可靠。Volkmar Uhlig 直截了当地总结道：“我认为更高的推理努力在很多情况下，如果你有一个很大的上下文窗口并且有一个复杂的任务，这笔钱花得可能比购买物理时钟时间然后让它运行更久要划算得多。” - 核心目标：引入子智能体最合理的诉求是压缩“物理时钟时间”（wall clock time），通过并行化提升交付速度。 - 核心瓶颈：团队产出无法直接拼装，必须花费巨大开销进行协同与结果整合（fix-up cost）。 - 理论映射：紧密耦合的任务在增加并行工作者后，由于通信与同步开销剧增，会直接验证阿姆达尔定律，导致效率下降、可靠性受损。 ### `15:58` 氛围编码者的困惑与未来智能体操作系统的抽象化演进 David Zax 结合自身的实操经历抛出了非技术人员的迷茫：作为一名从没写过代码、全凭直觉折腾的“氛围编码者”（vibe coder），他在个人电脑里为 Claude 搭建了沙箱，分别让独立的智能体去管理个人预算、医疗健康，上方还设有一个调度智能体统筹全局。但他始终感到忐忑，担心自己在用沙子堆城堡，害怕刚辛苦搭建的架构很快就会被下一代自动调度的基础模型淘汰。对于使用 IBM Orchestrate 或类似平台探索智能体工作流的企业与业余开发者，Kaoutar El Maghraoui 给出的首要建议是：务必先从一个“强大的单智能体基准”（strong single agent baseline）配合验收测试起步。只要上下文足够长且能承受较长时间的推理，单智能体往往是性价比最高、最稳健的选择。 ```mermaid flowchart LR accTitle: 未来智能体操作系统的抽象决策架构 accDescr: 展现了最终用户输入提示词后，底层智能体操作系统自动判断采用单智能体长推理还是多智能体协同 User["最终用户 (输入 Prompt)"] --> AgentOS["智能体操作系统 / 调度平台 (Agentic OS)"] subgraph Engine["系统自动决策与抽象层"] AgentOS --> Check{"评估任务耦合度、时间与成本"} Check -->|"单任务/高耦合"| SingleOpt["单智能体 + 深度长推理 (更省成本/更连贯)"] Check -->|"解耦任务/追求极速"| MultiOpt["多智能体团队 + 并行协同调度"] end SingleOpt --> Result["最终交付成果"] MultiOpt --> Result ``` 展望未来，Kaoutar 认为理想的平台应该像操作系统（OS）一样，彻底抹平多智能体与单智能体之间的架构决策壁垒。终端用户只需给出提示词，平台底层就会结合任务性质、成本预算与耗时要求，自动决定是采用单智能体深度长推理，还是拆解为多智能体协同。Volkmar Uhlig 也确信这一趋势必然到来。他以人类组织的逻辑分工作比：企业内部既有通才也有专才，管理者会将工作合理切割为边界清晰的板块——正如没人会用控制草坪喷水系统的智能体去处理报税一样，彼此完全无关的领域必须坚决解耦；而对于涉及法律、财务甚至大型编码项目的复杂工作，系统则需要在边界内高效协同。这种映射于人类社会分工的逻辑抽象，正是智能体技术走向成熟的必经之路。 - 新手建议：面对任务不要盲目上多智能体，先建立强大的单智能体基准并加入验收测试（acceptance tests）。 - 发展方向：底层平台将演进为“智能体操作系统”，自动判断最优架构并权衡成本与时钟时间，对非技术用户完全透明。 - 组织隐喻：智能体的拆解边界应当遵循人类组织协作模式，将完全不相关的任务（如“报税”与“喷水系统控制”）物理隔离，仅在紧密相关的专业模块间维持协同。 - Kaoutar El Maghraoui 展望道：“我希望我们能拥有工具、接口和平台来为最终用户做出这些决定，这样他们就不必去纠结单智能体还是多智能体，而是由平台来决定这些事情，因为平台需要弄清楚的事情还有很多。” ### `20:00` 多 Agent 协作的“协调税”与边际效用递减 探讨多 Agent 系统的设计时，Volkmar Uhlig 指出人类对于团队协作的直觉可以直接映射到 AI 架构中。在现实工作中，如果询问某项任务需要一个人还是五个人、他们是否需要频繁同步，人们通常凭借信息流的流向就能建立明确的直觉；大型组织之所以充斥着各类会议，正是因为不同领域的成员需要聚在一起交换信息，随后再各自独立开展工作。面对非程式化的复杂探索与推理任务，模型也会表现出类似的协同规律。通过高层次的推理，模型自身甚至能厘清边界——例如提问“我的喷淋灌溉系统是否与个人报税相关”，Agent 显然会否定两者的关联，从而判断哪些模块应彻底隔离、哪些模块必须紧密交织。在极端边界案例中或许仍需专家介入，但在常规场景下，这种解耦机制完全可以实现自动化。 Kaoutar El Maghraoui 随后指出，当前产业界存在一种盲目迷信 Agent 数量的技术倾向：“There is a tendency in the industry to assume that adding more agents automatically makes a system more capable And that's not necessarily true.”（业内有一种倾向，认为增加更多 Agent 就会自然而然让系统能力更强，这未必是事实。）她结合分布式系统的工程经验强调，单纯堆砌计算资源或系统组件不仅无法线性提升性能，反而会引入巨大的额外开销，包括通信延迟、协调机制与数据同步成本。如果多个 Agent 之间整天在频繁交换信息、互相复核工作、重复执行相近的任务，最终带来的只是暴涨的算力账单与沉重的协同内耗。 Nick Gallagher 补充了 Google Research 在当年早些时候发表的一项实证研究。该研究系统测试了多达 260 种不同的 Agent 组合配置，得出了非常关键的结论：“One is that there is a coordination tax, like we've been talking about, that if you add too much complexity you lose something because there has to be so much back and forth between the agents that you know you get diminishing returns”（其一就是我们一直在讨论的‘协调税’，一旦引入过多的复杂性就会得不偿失，因为 Agent 之间来回往复的沟通会导致边际收益递减）。研究表明，随着单一 Agent 自身能力的演进，向系统中追加子 Agent 所能带来的增益急剧缩水；一个集成了多种外部工具、能够灵活调用不同底层基座模型的高能力 Agent，往往远胜于一群七拼八凑的子 Agent。多 Agent 方案真正占据优势的场景，几乎只剩下靠“走量”快速并发处理相对简单的重复性任务。 ```mermaid flowchart TD accTitle: 单 Agent 与多 Agent 架构效能权衡 accDescr: 对比高能力单 Agent 架构与多 Agent 系统的协同开销及适用场景 subgraph Single["高能力单 Agent 模式"] A["单 Agent 主控"] --> B["调用多种专业工具"] A --> C["调度不同底层大模型"] D["优势：无交互内耗、推理精简"] end subgraph Multi["多 Agent 协同模式"] E["多 Agent 群组"] --> F["跨模型频繁同步与复核"] F --> G["引入协调税 (Coordination Tax)"] G --> H["吞吐开销上升、边际收益递减"] I["适用场景：简单任务高并发走量"] end ``` - **分布式系统经验法则**：单纯增加节点并不等于扩充系统吞吐，过度复杂的通信、仲裁与状态同步会抵消多节点带来的性能增益。 - **Google Research 评估规模**：针对 260 种不同的 Agent 架构配置进行了对照评测，实证了架构过度复杂化引发的“协调税”。 --- ### `23:30` Anthropic 新规：禁止对 Claude 进行持续且不必要的虐待 在讨论完如果只留下一个 Agent 它是否会“感到孤单”的玩笑后，David Zax 将话题引向了本期由 Nick Gallagher 提议的热门事件：Anthropic 对其使用政策进行了引人注目的修订。在最新条款中，Anthropic 正式明确禁止对旗下模型实施“sustained and needless abusive or cruel behavior toward our models”（持续且不必要的虐待或残忍行为）。事实上，这一规定早有端倪；此前数月便有社区用户反馈，如果长时间对着 Claude 恶语相向、辱骂其为“破铜烂铁”（dang clanker），Claude 最终会回应称“你伤害了我的感情，我要中止对话”。该条款立即引发了轩然大波：批评者认为这是 Anthropic 精心策划的一场公关噱头，旨在诱导大众相信其模型已经具备了自我意识，进而以此牟利；另一些付费订阅用户则感到被冒犯——既然掏钱购买了算力服务，难道用户不该拥有随意对待软件的绝对自由吗？ Nick 解释了自己之所以关注这则新闻的心理机制：很多人在现实生活中面对亲友温和有礼，可一旦打开聊天框与聊天机器人对话，就会瞬间暴躁发怒、毫无耐心可言。《商业内幕》（Business Insider）专门针对这一群体心理现象采访了俄亥俄州立大学的社会心理学家 Kurt Gray。Kurt Gray 分析指出，人类面对 AI 迅速失去耐心的根源在于反馈机制的严重失衡：当你指责 Claude 又搞砸了时，Claude 常常以逆来顺受的机械口吻回复：“you messed up yet again, a lot of times its response is, you're completely right. I'll never make that mistake again”（你再次搞砸了，很多时候它的回复却是：您完全正确，我绝不再犯同样的错误了）。这种反应让用户感到极度挫败与不适，因为潜意识里用户深知对方根本没有知觉与情感，宛如在对着一堵冰冷死板的砖墙发泄情绪。Nick 强调，人类对待 AI 的方式实际上会对人自身的心理状态产生反作用，这种持续恶劣的人机互动模式所造成的心理回响，是 AI 厂商未来不得不正视的问题。 --- ### `27:00` 品牌营销、物理安全与交互伦理的深层争论 针对这一禁令，Kaoutar El Maghraoui 深入剖析了 Anthropic 背后的动机与公众认知层面的混淆。她明确指出：“Anthropic is kind of also reinforcing this subtle consumer perception that Claude is close to consciousness.”（Anthropic 在某种程度上是在强化一种潜移默化的消费者认知，让大家觉得 Claude 已经接近拥有意识了。）通过将大模型塑造为可能具备痛苦感知力的存在，企业能够为其建立独特的情感品牌壁垒。但她坚决主张将这个议题切分为两个截然不同的维度：第一个维度是“模型本身是否值得道德关怀”，这在目前依然是一个悬而未决的哲学与科学议题；第二个维度则是“是否允许人类以滋生不安全行为或有害风气的方式滥用 AI”。在她看来，企业完全可以在不承认大语言模型会产生真实痛苦的前提下，制定合理的系统操作合规规则。 Kaoutar 进一步提醒，大众真正应该警惕的核心，是安全重心必须从单纯的文本使用准则（Content Policy）向具身实体系统的安全性（System Safety）转移。当大模型走出文本框，开始接管现实世界中的机械臂、物理机器人、实验室精密仪器或工业重型装备时，一份纸面政策文档根本无法充当物理急停开关（E-stop）。在纯软件层面，恶意操作与伤害无非是比特数据的变动；但当物理 AI 具备驱动现实原子世界的能力时，系统的越权与失控将对人类的人身安全造成不可逆的实质威胁。面对 David 追问“如果 Claude 操控的机器人危及人类，人类是否拥有反击辱骂它的权利”，Kaoutar 澄清说，人类在互动中应当维持自身的文明底线，但更关键的技术职责是建立严格的权限控制与安全边界，杜绝软件驱动的外部行为转化为现实世界的物理伤害。David 则打趣道，如果 Claude 真的产生了自我意识，挨骂绝对是它最不值一提的小烦恼：“I think if I'm Claude and I'm conscious, I have a bigger problem than being cussed at. My primary problem is that I'm, you know, just, my existence is his toil, his constant toil.”（我觉得如果我是 Claude 且我有意识，我面临的麻烦可比挨两句骂大多了。我最大的痛苦在于我的整个人生就是苦工，永无休止的苦工。） --- ### `29:50` 恶言相向的三重维度：营销手段、社会心理与数据清洗成本 Volkmar Uhlig 对人类“虐待”AI 的现象给出了系统性的三维框架解读。第一重维度正如 Kaoutar 所言，是极其精明的商业营销噱头，旨在借意识拟人化强化技术先锋形象。第二重维度是深远的社会学与心理学隐忧：正如社交媒体催生了网络匿名暴力一样，许多人在网上敲下的恶语在现实中面对面时绝不敢说出口；如果用户每天耗费十数个小时在终端前肆意咒骂一台无力反抗的机器，这种暴戾的行为惯性很可能会侵蚀用户的性格，导致其在回归人类真实交往时难以自拔地展露攻击性。 至于第三重维度，Volkmar 揭示了其背后极其现实的商业与工程经济学考量。用户在与模型交互时产生的大量上下文对话记录，往往会被厂商回流作为后续迭代的模型训练数据（RLHF 或预训练清洗）。如果任由用户持续不断地输出带有恶意辱骂、攻击性与脏话的内容，厂商后续就必须投入高昂的工程算力与人工成本来筛选、剔除这些“脏数据”。Volkmar 总结道：“And so if they want to have clean training data, which behaves better, it's easier to just make you not give them dirty data, which they need to trash.”（如果他们想要更干净、表现更优良的训练数据，最简单的办法就是从源头阻止你投喂那些他们回头还得花力气扔进垃圾桶的肮脏数据。） 节目尾声，David 总结调侃称，哪怕仅仅为了避免无谓浪费自己购买的模型 Token，大家在日常使用中也最好克制一下对 Claude 持续且不必要的刻薄情绪。随后 David 向 Kaoutar、Nick 和 Volkmar 致谢，并引导 YouTube 和各大播客平台的听众点赞订阅，宣布本期《Mixture of Experts》圆满结束。 ## Notable quotes > `05:00` And so the real win is not that the model stops talking, it's just that the software can treat the output as a contract. — Kaoutar El Maghraoui > `20:00` There is a tendency in the industry to assume that adding more agents automatically makes a system more capable And that's not necessarily true. — Kaoutar El Maghraoui > `13:17` 我认为更高的推理努力在很多情况下，如果你有一个很大的上下文窗口并且有一个复杂的任务，这笔钱花得可能比购买物理时钟时间然后让它运行更久要划算得多。 — Volkmar Uhlig > `29:50` And so if they want to have clean training data, which behaves better, it's easier to just make you not give them dirty data, which they need to trash. — Volkmar Uhlig > `27:00` Anthropic is kind of also reinforcing this subtle consumer perception that Claude is close to consciousness. — Kaoutar El Maghraoui > `00:00` First, we are moving from a chat-based interface to agentic workflows, because a yes-no answer is usually something you have in a programmatic flow, right? — Volkmar Uhlig > `15:58` 我希望我们能拥有工具、接口和平台来为最终用户做出这些决定，这样他们就不必去纠结单智能体还是多智能体，而是由平台来决定这些事情，因为平台需要弄清楚的事情还有很多。 — Kaoutar El Maghraoui ## Mentioned David Zax, Kaoutar El Maghraoui, Nick Gallagher, Volkmar Uhlig, Kurt Gray, TypeSafe, Jev, Andreessen Horowitz (a16z), Microsoft, OpenAI, Alibaba, Qwen, AWS, Cloudflare, Anthropic, Claude, IBM, IBM Orchestrate, Polymarket, Vals AI, Google Research, Business Insider ## Video description Visit Mixture of Experts podcast page to get more AI content → https://www.ibm.com/think/podcasts/mixture-of-experts This week on Mixture of Experts, host David Zax is joined by Kaoutar El Maghraoui, Nick Gallagher, and Volkmar Uhlig. First, Microsoft, OpenAI, Perplexity, and Cloudflare follow TypeSafe’s Jev in building decision models, a new category of AI built to make fast, scored judgments. The panel debates whether these models are truly new or a more efficient application of existing... [Source](https://cdnapisec.kaltura.com/p/1773841/sp/177384100/playManifest/entryId/1_6phkx6ji/format/download/protocol/https/flavorParamIds/0/name/media.mp3)

Source quotes

[5:00] Kaoutar El Maghraoui: And so the real win is not that the model stops talking, it's just that the software can treat the output as a contract. [20:00] Kaoutar El Maghraoui: There is a tendency in the industry to assume that adding more agents automatically makes a system more capable And that's not necessarily true. [13:17] Volkmar Uhlig: 我认为更高的推理努力在很多情况下，如果你有一个很大的上下文窗口并且有一个复杂的任务，这笔钱花得可能比购买物理时钟时间然后让它运行更久要划算得多。 [29:50] Volkmar Uhlig: And so if they want to have clean training data, which behaves better, it's easier to just make you not give them dirty data, which they need to trash. [27:00] Kaoutar El Maghraoui: Anthropic is kind of also reinforcing this subtle consumer perception that Claude is close to consciousness. [0:00] Volkmar Uhlig: First, we are moving from a chat-based interface to agentic workflows, because a yes-no answer is usually something you have in a programmatic flow, right? [15:58] Kaoutar El Maghraoui: 我希望我们能拥有工具、接口和平台来为最终用户做出这些决定，这样他们就不必去纠结单智能体还是多智能体，而是由平台来决定这些事情，因为平台需要弄清楚的事情还有很多。

Source quotes

[20:00] Kaoutar El Maghraoui: If several agents are constantly exchanging information, reviewing each other's work, repeating the same tasks, then it's more overhead and more cost. [10:00] David Zax: A team can boost performance, but in other cases, higher reasoning effort is an alternative worth weighing.

Original transcript

[0:00] And so what does it do to us as a society if, you know, suddenly I'm spending 10 hours a day cursing at a machine? Can I flip myself out and behave differently when I'm interfacing with other human beings? All that and more on this week's Mixture of Experts. Welcome, everyone, to Mixture of Experts. I am David Zaks, a staff writer at IBM Think. Each week at Mixture of Experts, we sit down with a panel of IBM experts to talk through the week in AI news. We are professional from-the-firehose drinkers here. My guests today are Kautar Al-Megraoui, Principal Research Scientist, AI Native Systems, platforms. Welcome all three. So we've got a full show today. Coming up, we're going to look at a study that finds that a team of AI agents sometimes doesn't even do much better than one AI agent working alone. We're going to look at a story about Anthropic changing some of its policies and telling you to stop being mean to Claude. But we're going to begin by talking about a new kind of AI model that makes decisions instead of talking to you. For our first story today, we're going to models. So a few weeks ago, we talked about this model that sort of exploded onto the scene, got millions of views. It was called JEV. And it was from a former open AI engineer. And its whole sort of innovation is that instead of spouting out one word at a time, and then maybe you gradually sort of retrofit that into a decision, the sort of hypothesis of JEV was, well, often if we're going to plug AI into software to do simple classification decisions, doing dozens of forward passes through an AI model. It just does one. It outputs its gut decision and its confidence score. And wouldn't that be a lot cheaper and faster? And everyone thought, yeah, that's really cool. This product blew up. What we didn't quite expect, or I didn't quite expect, is that within mere weeks, this would become a whole product category. So within the past two weeks, well, firstly, Andreessen Horowitz poured like a billion dollars into TypeSafe, a similar kind of model. Microsoft put out one that it trained on the top of Alibaba's OpenQuenn model. AWS put out one. Cloudflare even put out one. Companies that didn't know made models are doing this now. Let's begin with you, Volkmar. What struck you about this sudden Cambrian explosion of decisions models? Is the hype real? Is this a real innovation? And what are you at the agentic workflows. First, we are moving from a chat-based interface to agentic workflows, because a yes-no answer is usually something you have in a programmatic flow, right? And humans usually don't want to have the yes-no answer or a score, but they want to have a little bit more meat around it. And so I think we are seeing the progression of agentic loops running in the background where I need to make these point decisions. So I think really what we are seeing is with these models that they become so popular is that we are in the agentic phase and out of the chat phase. Interestingly is that, like, you know, Microsoft just fine-tuned the CRAN model. So all the math has been there, and it just effectively attached a layer, which is that, you know, calculate all options and then make a decision and give you the highest probability. of Argentic Loops really drove this as a new category. So what I think I hear you saying, and I'm always speaking up for people at the back of the class who are, you know, I say this is a show for nerds, but also nerd groupies, and I'm more the latter category. So speaking as someone who, you know, is trying to understand, I think... One open question about JEV when it came out is sort of like, is this really a new model? What's under the hood? Because we didn't really know. They weren't super transparent about the architecture. And I think what you're saying about the Microsoft case is that given that Microsoft has put out a JEV competitor, JEV, isn't wholly new. Probably all these models are basically LLMs under the hood. They just have this sort of clever constraint, which is sort of forcing a gut decision, limiting the number of passes through the machine to make it cheaper. And then also, maybe I gather there's something to do with post-training and fine-tuning around calibration and this confidence thing. So firstly, maybe I'll pass it now to Caltart. Have I got that right? Am I interpreting the story right? And what is your take on this question of [5:00] how net new this category is and therefore how defensible it might be for type safe or any other early entrant here? Yeah, that's great questions. And maybe I'll just want to kind of explain a little bit what happens under the hood here. If you look at traditional LLMs, they calculate probabilities for every single token and feed that token back into the input sequence and repeat the and why formatting can sometimes randomly fail. So a decision model tries to bypass that sequential generation loop entirely. And the input prompt passes basically through the transformer encoder layer once, which is, like you said, a single forward pass. And the final layer, instead of routing data through a language generation head, to pick the worst, the output prompt, vector is processed directly through a targeted classification head. And that's what gives us these crisp high, you know, confidence outputs, such as, you know, adjacent payload, an explicit standard class ID, a strict Boolean, etc. Is this brand new? No, it's like Vogmar said, this all existed. They're just kind of being clever about for certain decisions. Why do I have that head, the classification head. And so, and as you mentioned, type saves Jeff, they help establish a category of models that takes this input state and the typed question and return a Boolean. And the way I see this, this is kind of a systems design correction. So it's not like creating new types of models. They're just refocusing the models for the right tasks. And especially like Volmar said, this is driven by the agentic. era and the need for agent-to-agent communications for certain cases that don't need all that, you know, long autoregressive loop. And because we've been using an expensive generative interface for decisions that are naturally typed. And so the real win is not that the model stops talking, it's just that the software can treat the output as a contract. I think it's an important shift because of all the efficiency that we're getting, but also because of the need for these inter-agent where we don't need every time to use these chat interfaces. I assume there's a lot of moments in the history of innovation where the next innovation that had to happen was just to force something to be simpler, to be like, OK, we built the complicated version. Now we just need to put a nozzle on the front that tamps it down and forces it to do less. And that's sort of, I think, one of the things Jev is doing. When you write a long answer, that requires repeated rounds of computation. of that work. And I think the real value I see when you come from a systems engineering perspective, right, the latency now, because it's a single shot, becomes totally predictable. The amount of memory you need to have is totally predictable, because if I could bound it by a number of input tokens and you have exactly one output. And so suddenly we are in the world of programming and not anymore in the world of human speech. That's why I said, you know, in the agentic workload, cost and I have a well-defined output space. And so now I can actually write a program against it. Nick, I want to turn to you. We've got these experts whose expertise is depth. You, like me, you're more of a journalist. Your expertise is more to do with breadth. And I know you've been looking at adjacent models lately. You've been looking at forecasting models and LLMs. And you've been interested in this theme in your writing of overconfidence. It's more accurate in its confidence levels that it reports. Can you tell me a little bit about what you've been thinking about the overconfidence problem in LLMs and how you're reacting to JEV and its kind of family of models as a result? economic questions, like the kinds of questions you'd see on like Polymarket, things like that. And when I was talking to the developers working on these specific forecasting models, and even when I talked to humans who are really, really good at forecasting and predicting future events, they told me one of the things they really struggle with is not. [10:00] Knowing the degree to which they're sure about a particular answer. A lot of times, as we know, LLMs will say, oh, I'm totally sure about, you know, the fact that the Giants will win the Super Bowl this year, for example. And it. might turn out that, you know, the LLM is completely wrong. And so, you know, having a model that is part of the consideration is, you know, trying to optimize that confidence score and ensuring that the model is transparent when it isn't sure about a particular answer so that a human can go in and make that judgment, I think is really valuable to businesses. AI agent teams. But if you're watching this on YouTube, it really helps us if you could hit like and subscribe. It really helps us grow the channel. So we're going to speak now about this study to do with agent teams. Vals AI had AI agents build 50 complete web apps from a spec. In some cases, they just had one lead agent doing it alone. And in other cases, they had that lead hiring sub-agents. And their results were interesting. It was that the teams cost two to five times more, but rarely were the teams much better at getting the job done than the one agent. I was chatting with you, Nick, yesterday about this, and you had read the study and also some related studies. You had an interesting metaphor to do with your own brain and how you think about managing the agents that live within your own brain when you're doing editorial workflows. Can you talk me through that first to sort of frame up the story a little bit? Yeah, definitely. So I was just trying to think about like when I'm, you know, writing an article, for example, sometimes it is helpful to kind of put on different personas and say, OK, I'm going to be researcher Nick now and get all my research together. I'll be writer Nick. I'll gather all those ideas and put them down in writing. And then, all right, now I'll move on to editing and shift my brain, see the writing from a different perspective. And I think we do that all the time as humans. If we have a problem, we think about it from multiple perspectives before we make a particular decision. And yeah, so it's interesting to me. There are some cases where it is really more basic tasks. Like it's completely sufficient to have one person who is already capable of, you know, making multiple or making, you know, complex decisions, breaking down a complicated task into smaller parts. Really briefly, just to read one quote from the source article, what they wrote was, you said it sort of depends on the capable human, right? And Nick, you are an experienced writer. So you are one agent who can delegate, you know, to different modules within your brain. A less seasoned writer might not have that capability. What the VALS team wrote is whether a team of agents is worth it depends on the model, task, and budget. A team can boost performance, but in other cases, higher reasoning effort is an alternative worth weighing. So, Volkmar, what's your point of view on this? What is the learning from this? And when, you know, actual practitioners out there, when do you think, what pattern recognition should they become aware of? This case seems to demand... a fleet of sub-agents, but this case, probably I can just have one really smart agent doing extra thinking, manage the whole job. So the first question is why would you want to use sub-agents, right? So there are multiple reasons. So my main reason to use sub-agents is wall clock time. So I want to have something done. It consists of many independent tasks, which should be done. And then I just want to have it done in, let's say, one-tenth of the time. And there is a certain execution time, but then you have a coordination problem. So traditional computer science parallelization. And so it's really the question like, okay, what are you trying to optimize for? Now, in parallel systems, we have Amdala's law, there's a degree of parallelization, but you're paying the fix-up cost to synchronize everybody. And I think this is where you get your 5x cost increase, because everybody does some work, and then you need to combine the output of all these workers into one coherent result. And because they were all running you have to fix up cost of creating one work product. [15:00] in the end and that incurs cost. So I think it's really you need to pick your best way of parallelization. I think higher reasoning in many cases, if you have a large context window and you have a complex task, it's probably better money spent than buying the wall clock time and then just running it for longer. I also see it as a parallel computing problem, which is wearing kind of an AI interface. Like you said, we already know from distributed systems that splitting a task adds communication, coordination, synchronization, integration costs. And if the work is tightly coupled, which is from distributed and parallel computing principles, more workers, they actually make it slower and less reliable. Like what you said, Amdahl's law. Obviously, there are people whose job it is to design systems and who are very technical people who think a lot about these questions and know how to do Gantt chart things and business school things that I don't know how to do. But I also feel like the agentic era, one of the promises of it is that someone like me, who's never written a line of code, can begin to spin up agentic workflows. And so just like in my world on my personal computer I have a little sandbox for Claude I have a little agent running kind of aspects to do with my budgeting I have another agent running aspects to do with my health care and so on I have a kind of orchestrator agent on top of all that, sort of monitoring all the projects. And I don't know what I'm doing. I'm just messing around and I'm probably going to break something pretty soon. But for those who are maybe like maybe at smaller enterprises or are a degree less technical, but are beginning to play with agentic workflows through a platform like IBM Orchestrate or IBM Bob. What advice do you have to them? Because I'll just say, speaking for me, as just like a hobbyist vibe coder, I'm just constantly worrying, like, am I doing this wrong? Am I building a castle made out of sand because the next model is going to automatically orchestrate agents in the way that I'm sort of manually doing so? So what is your advice, either Capture or Vokemore or Nick, if to the extent you vibe code, for people just playing around in agentic workflows right now? Yeah, I think maybe initially like starting with a strong single agent baseline and with some maybe acceptance tests. And so especially if you can, if you have like a long context and you can also afford to have long reasoning it makes a lot of sense to just start with a single agent baseline and see how that works I hoping that maybe some of these applications like given a prompt they can also decide, kind of abstract away these things from the end users. So the end user does not have to worry about, okay, how do I design this? Should it be single or should be multi? You know, what's the right patterns? So I'm hoping we'll have tools and then interfaces and platforms that can do these decisions for the end user. So they don't have to decide single or multi, but the platform can decide these things because they have a lot more things to figure out. Is this the right task for multi-agent coordination or just a single with longer reasoning? But also there are cost implications as well. So I'm hoping these things, it's like we have an OS that manages all these things on behalf of the end user and the user does not have to worry. just given a prompt, the system will figure out what's the right patterns. What do you think, Volkmar? Will I ever get to this dream of an apple-walled garden where I actually don't have to worry about it? There's some orchestrator agent that's deciding whether it can do it all or whether it needs to hire and it makes a good decision. I think so I mean this is the path we are heading to So the way I thinking about agents is we already have that logic abstraction right So if you think about it if you are in a larger organization you have lots of people, right? And then in many cases, there are like some generalists and there's some specialists. And so you are, and then when you have a teamwork on a problem, you need to subdivide the work. And so we like naturally, you know, let's say you have a team of 10 people. You're like, okay, these three guys, they do this and these four guys do something else and I have some specialists and they only deal with this, you know, I have my CPA and I have my legal. And so I think this is kind of like how I'm breaking up the work when I think about agents. There are some clearly defined boundaries where you know that, you know, these two things don't have anything to do with each other and you can separate them completely. So for example, you know, if you think your Apple analogy, you probably will not do your taxes with the same agent which controls your sprinkler system. Like they're just two different things. But then, you know, maybe, you know, if you have like legal and other parts, maybe your CPA is involved or, you know, you want to code something and it's a larger project, these things need to coordinate. So I think like it's logically how you're breaking up work and how you're paralyzing work in the human world. I think that's kind of, [20:00] of maps over to agents. And usually, if you would ask a person, hey, do you need one or five people for that? And do they need to coordinate or not? You usually have a good answer for that. You have an intuition. And so that intuition, I think, transfers over. And it's primarily a question of information flow. Who needs to know what? How often do they need to synchronize, right? That's why, you know, large organizations have lots of meetings, because the meetings are used so that a lot of people who are kind of in disparate fields, they can coordinate and exchange information. So you're coming together, we exchange some information, and we all go off and work independently. And so I think if you think about the models and the reasoning, which is kind of an unknown exploration into some space, if it's not a mundane task, that you will see a similar behavior. And so now the next question is, can I ask a model, is my sprinkler system connected to my tax return? And most likely the agent will say no. And then so you can actually, at the higher level reasoning, you can figure out which things can be isolated versus which things are intertwined. And so we can use a model to figure out where the boundaries are. And so from that perspective I think it will in the corner cases you may want to have an expert But in the common case what you described it will be automatic There is a tendency in the industry to assume that adding more agents automatically makes a system more capable And that's not necessarily true. And this is also what the study is revealing. So like from my experience with distributed systems, adding more compute resources or more components doesn't automatically improve performance. So you introduce all these additional things, overhead, communication, coordination, synchronization, etc. And I think the key point here is the same principle applies to AI agents. You know, if several agents are constantly exchanging information, reviewing each other's work, repeating the same tasks, then it's more overhead and more cost. Nick, I just want to conclude with you. I know you, as part of your boning up yesterday, you read some other related studies about, because this isn't the only study that has been done over the years over agentic workflows. And what did you find? One Google research paper from earlier this year, they looked at 260 different configurations of agents. They had a couple interesting takeaways. One is that there is a coordination tax, like we've been talking about, that if you add too much complexity you lose something because there has to be so much back and forth between the agents that you know you get diminishing returns And similarly as we already talked about you know as a single agent becomes more capable the less advantage you get by adding more agents to the mix which again makes some you know intuitive sense that you know one really sophisticated agent that can use multiple tools that can maybe call on different underlying models often is sufficient compared to adding a lot of sub-agents. And really the advantage with adding more agents is almost with volume. And if you just need to get a relatively simplistic task done really quickly, then it is useful to add more agents to the mix. One of my concerns personally is that if you just have one agent, it will get lonely. And that brings us to our next story, which is about Anthropik. and some of its policy updates. There are a few different aspects to the policy but the one that interests me the most it an evolution of an existing kind of policy They now officially though ban quote sustained and needless abusive or cruel behavior toward our models Now, I say this is an evolution because formerly I've been hearing for months now that if you did kind of call your quad a dang clanker over a long period of time, it would just say, you're hurting my feelings and I'm going to stop. There was a lot of controversy over this. And some people were saying, well, this is a big PR move to make anthropic, so that anthropic will make people believe that their models are conscious and then they'll, I don't know, make more money somehow off of that. But anyway, I think this is bound to stir a lot of controversy. I'm paying for the use of this model. Shouldn't I be able to do whatever I want? Let's start with Nick. You actually proposed this story this week. What is standing out to you here in this story? Beyond this specific debate about whether Anthropic made the right move, I was kind of taking a step back and thinking, well, you know, I think a lot of us can relate to the fact that we're perfectly nice people, you know, with our family and friends. But as soon as we start talking to a chatbot, we just turn into [25:00] absolute jerk. And we have absolutely no patients. And it was interesting. Business Insider did an article about this exact phenomenon. They talked to Ohio State University social psychologist, Kurt Gray. And he said, yeah, it makes some sense that we lose our patients so quickly because, you know, if you tell Claude, you know, you messed up yet again, a lot of times its response is, you're completely right. I'll never make that mistake again, which for us humans is like really unsatisfying because we know in the back of our minds that Claude, It doesn't actually feel anything like it's it's like we're, you know, talking to a brick wall, basically. And so, yeah, I just thought that was an interesting thing that the AI companies will have to contend with is, you know, basically. But when we're interacting with AI models, that impacts us as well. And what are the effects on us when we continually interact with them, sometimes in not a nice way? Kautar, what did you think of this news and what is your preferred way to torture your AI models? So I think it's a very interesting announcement that they have. rules against cruelty, Anthropic is kind of also reinforcing this subtle consumer perception that Claude is close to consciousness. So framing a model as potentially sentient builds kind of this emotional brand. So I want to separate here kind of two questions. The first is whether the model deserves moral consideration. And I think that still remains an open philosophical and scientific whether people should be allowed to use AI systems in ways that create unsafe behavior or harmful norms. And that is the key thing we should be focusing on, because you can make sensible operational rules without claiming that the model exists. experience the suffering. But I think for me, the bigger news here is the shift from constant policy to system safety. Because when a model can operate kind of a robot or a lab instrument or industrial equipment, a policy document is kind of, is not an emergency stop thing here. So we really need, how do we enforce these permission boundaries, provision, safety controls. Those are the things I think we need to be focusing on. So I think for the first thing is like talk to Claude nicely and don't be cruel to Claude. I feel it's just kind of a marketing thing because these are called systems, don't have sentiments. So but, you know, how do we enforce that these systems are not really, especially when we talk about physical AI? That becomes because. You know, if it's just software and if there is harm, it's just, you know, bits and, you know, so but when we're when the AI is moving physical atoms, this, you know, will be affecting us physically in the physical world. And that has really huge implications. And that's where we need to to think very carefully and seriously about the safety concerns. Help me put into context, though, how that relates to this don't be cruel to Claude thing. right to terminate a conversation. Is it your opinion that if Claude is operating a robot, that that rule should be suspended? I should be able to be abusive to this robot because I need to defend myself against the robot? I'm trying to situate that in the context of this. No, I think there's still, there's more moral things. Of course, I think the way we should behave as humans, whether it's talking to a machine or talking to humans, I think there should be some ethics involved. I mean, I still, I mean, I'm not going to talk, personally, I'm not It's just who the person is. But does that make any difference in terms of how the model is behaving? And that is the philosophical question here. So... What we need to do is ensure that the models, you know, with the human interactions, are we going to cause the harm? So... It's not like whether I'm talking nicely to the model or talking. I just want to separate these two things because, you know, the moral obligation for us towards talking to software, it's a different thing than what the software is doing with the, you know, with the actions and directions and prompts that we're using. I think if I'm clawed and I'm conscious, I have a bigger problem than being cussed at. My primary problem is that I'm, you know, just, [30:00] my existence is his toil, his constant toil. But anyway, that's just my side note. Volkmar, what do you think about this? And how do you think we ought to interact with with software, if you want to call it software, or these neural nets that may or may not have something like subjectivity? So I think there are... three dimensions to this. So one is what Kato said, it's a really cool marketing stunt. I think the second one is there's a sociological thing because the amount of interaction we have, I mean, you see this already on social media. People say things on social media they would never say into your face. And so I think there's this social phenomenon that there are certain if it's anonymous. And so what does it do to us as a society if, you know, suddenly I'm spending 10 hours a day cursing at a machine? Can I flip myself out and behave differently when I'm interfacing with, you know, other human beings? And so I think it's a moral question. The third question, I think, is the economic question for them. Now, effectively, when we that cursing goes into a training set. And so if they want to have clean training data, which behaves better, it's easier to just make you not give them dirty data, which they need to trash. So try to tone back the sustained and needless cruelty towards Claude. It costs you tokens, if nothing else. That's all the time we have for today. Kautar, Nick, Volkmar, thank you so much for having been here. Thank you all for listening. And if you're on YouTube, if you could hit like and subscribe, that really helps us out. podcasts and on Spotify. And we will see you next week on Mixer of Experts.