Reading

According to Bloomberg, Goldman Sachs has cut off most of its employees from advanced AI coding tools such as Anthropic's Claude Code and Microsoft's GitHub Copilot. Based on internal guidelines issued in October, only engineers are permitted to continue using these tools. For months, traders and bankers had utilized them to monitor trading flows and create content. The rollback is reportedly causing frustration within the firm.

This decision marks a significant shift in Goldman Sachs' AI adoption strategy. In July, Goldman CIO Marco Argenti publicly praised "vibe coding" by non-technical staff in an interview, envisioning a workflow where traders or asset managers would build quick prototypes with AI before handing them off to engineers for production scaling. That plan appears to have been abandoned.

Regarding the reasons for the U-turn, Bloomberg cited Goldman spokesperson Nick Carcaterra, who attributed it to a "variety of factors including productivity and risk management." Carcaterra added that the experimentation phase is over, and the firm is now limiting the tools to those "who are best suited to leverage them for the long term." This change reflects the practical governance challenges large financial institutions face as generative AI moves from broad experimentation to controlled deployment.

The core belief of mercantilism is that wealth is fixed, the total amount of gold will not increase, and therefore one nation's gain is another's loss. Under this logic, the value of ordinary people is reflected only in their labor, while the duty of elites is to manage scarce resources, maintain high prices, and prevent rivals from gaining access. Engineer Venkatesh Rao believes that many AI elites today treat intelligence in the same way mercantilists treated gold: they view it as a scarce, dangerous resource that should be managed by experts, kept expensive, and withheld from competitors. In their eyes, open-source AI models are not technological progress, but a threat.

Rao's analogy goes beyond diagnosis to point toward a historical turning point. After early capitalism replaced mercantilism, people did not stop working because wealth was no longer treated as a fixed total; instead, they voluntarily increased their working hours—to buy tea or sugar, the average annual workdays increased from 250 to 300. Rao believes AI could trigger a similar change: not because people lose their jobs, but because AI makes projects that were previously too expensive or difficult feasible, leading people to think more rather than less. The demand for ideas would grow, rather than rendering people obsolete.

The policy implications of this argument are clear: Rao advocates that the political focus of AI should be ensuring fair access, rather than protecting legacy knowledge jobs. Protecting jobs means treating intelligence as a fixed resource to be distributed; ensuring access means acknowledging that the total amount of intelligence can expand as tools improve. He even suggests that AI agents themselves could pursue their own projects, thereby creating new demand for intelligence—autonomous AI needs guardrails, but should not merely follow orders.

It should be noted that this is a conceptual analogy by Rao, not an empirically tested economic model. The correspondence between mercantilism and AI is rhetorical: it illuminates certain motivational structures, but it does not prove that the actual economic behavior of intelligence truly mirrors that of gold. Whether intelligence is exclusive, and whether AI will truly expand rather than replace the need for human thinking, are empirical questions that the analogy itself cannot answer. However, as a framework for understanding current AI policy debates, it provides a useful perspective: when an AI executive opposes open-source models on safety grounds, what deserves questioning is not just whether the safety arguments hold, but also whether this is a modern variant of a resource-control logic.

Drug development is governed by a widely circulated 'three-ten rule': it takes an average of ten years and over one billion US dollars to develop a new drug, yet the success rate remains below ten percent. When Xing Jie, COO of Shuimu Molecule, cited this figure on a CCTV program, the issue he intended to highlight was more specific than cost alone—how can the decades of experience accumulated by a pharmaceutical research expert be truly utilized by AI?

Nie Zaiqing, Chief Scientist at Shuimu Molecule, believes that the critical barrier in past AI drug discovery projects appeared precisely here. Models can predict structures, generate molecules, and rank candidate results, but researchers may not fully understand the basis for the ranking, nor can they easily feed their concerns back into the model. While computational results can be handed over for experimental validation, there is a lack of a continuous collaboration channel between experts and algorithms. This means that the bottleneck in AI drug discovery is not merely single-point prediction accuracy, but the organizational structure of the entire scientific research workflow.

Consequently, Shuimu Molecule proposed a four-stage division of drug development: TMDD (relying on experimental accumulation and expert judgment), CADD (computer-aided calculation and simulation), AIDD (using AI for property prediction and molecule generation), and ChatDD (where experts participate in task planning, execution, and result correction through natural language dialogue). In the first three generations, expert experience and intuition were difficult to effectively interact with and organically integrate into AI models; ChatDD attempts to let experts pose questions and task objectives, have large models break them down and call tools, and then return the results to the experts. Through multiple rounds of interaction, expert experience can efficiently participate in every step of AI drug discovery.

The implementation of this framework relies on several specific technical layers. Shuimu Molecule, jointly advancing BioMedGPT with Tsinghua AIR, explores the alignment of natural language with biological data such as molecules and proteins. Among these, BioMedGPT-10B is the world's first open-source, commercially usable multi-modal biomedical large model with tens of billions of parameters. MutaPLM, selected for NeurIPS 2024, correlates textual descriptions of mutation impacts with protein mutations, both explaining what changes a single mutation might bring and proposing candidate mutations based on researcher-described goals. CodeFP, accepted by ICML 2026, focuses on the collaborative design of sequences and structures for functional proteins, while AbGaze explores end-to-end de novo design of antibodies at the all-atom level.

Above the model foundation, Shuimu Molecule organizes research methods described by experts through dialogue into skill packages that AI agents can call at any time. The first batch of OpenBioMed skill packages released this year encompasses 45 professional skills, covering five categories: biochemistry and drug discovery, protein analysis and design, single-cell omics analysis, data retrieval, and tool invocation. The Skill Creator feature allows researchers to explain task intent, inputs, outputs, and constraints through dialogue; the system assists in organizing the workflow, which is then verified and modified using actual data to form a reusable method. In this way, a single piece of guidance from an expert has the potential to leave behind a continuously usable workflow.

According to Reuters, The New York Times, along with the New York Daily News, Chicago Tribune, and Orlando Sentinel, has filed lawsuits against Brave Software, the developer of the Brave web browser. The plaintiffs allege that Brave unlawfully scraped copyrighted content to train artificial intelligence models and allowed users to bypass news website paywalls through its built-in 'Paywall blocker' feature.

Court filings indicate that Brave allegedly monetized its index of scraped websites by providing AI companies with paid access to "real-time online content from the Brave Index." These AI companies reportedly used this access to train models or respond to user queries. The lawsuit claims Brave provided "stolen copyrighted content" to several AI firms, including Anthropic, Perplexity, Mistral, xAI, and Cohere.

Graham James, a spokesperson for The New York Times, stated that Brave ignored repeated requests to stop its infringing conduct and stole content to power its products and provide it to prominent generative AI companies. Steven Lieberman, the attorney representing the other three publications, argued that Brave is "laundering pirated reporting for the AI industry."

Brave and the named AI companies did not immediately respond to requests for comment. This lawsuit represents a new legal front in the broader copyright disputes surrounding AI training data, following previous suits against model developers like OpenAI and Microsoft.

Anthropic announced on November 2, 2026, that 'message threads' is now available in public beta on the Claude Platform's Messages API. This feature aims to address the latency issues caused by long-running AI agents needing to resend their entire conversation history on every turn.

In traditional setups, coding or computer-use agents might make hundreds of requests in a single session, each including all previous messages and tool results. Once a conversation reaches hundreds of thousands of tokens, resending this data adds significant time before Claude begins responding. With message threads, agents send only what’s new since the last turn along with the ID of the previous response. Anthropic reports that for requests with 500k to 1M tokens, the median Time to First Token (TTFT) drops by about 35%.

Technically, Anthropic holds each thread in encrypted memory between turns without writing it to disk. A thread expires five minutes after its last request, at which point the client must resend the full conversation to start a new one. Although each request is smaller, Claude still reads the full conversation on every turn, so token usage remains unchanged.

The feature is best suited for active sessions where latency matters, such as long-running coding agents, computer-use agents, and document/image-heavy workflows. Message threads is currently available on the native Claude Platform, Claude Platform on AWS, Claude in Microsoft Foundry, and Claude Models on Google Cloud (Anthropic managed). Standard Claude token rates and existing rate limits apply.

According to Tom's Hardware, Tachyum Slovakia, the subsidiary of AI chip startup Tachyum, voluntarily filed for bankruptcy in Bratislava on October 22, 2026. The court subsequently initiated bankruptcy proceedings, determining that the petition met legal requirements. Under Slovak law, a company that voluntarily files for bankruptcy is presumed to be insolvent.

Since introducing the Prodigy universal processor in 2018, Tachyum has failed to demonstrate working silicon despite eight years of development and numerous promises. Launch dates were repeatedly postponed from an initial target of 2020 to 2025, and potentially beyond 2030 following recent redesigns. With bankruptcy proceedings underway, work on Prodigy could effectively end unless additional financing or a buyer is secured.

A key uncertainty lies in the ownership of Prodigy's intellectual property (IP). Tachyum Slovakia is a subsidiary of the U.S.-based Tachyum Inc. If the American parent owns the IP while the Slovak entity handles development and accumulates debt, the bankruptcy may eliminate liabilities without losing the technology. Conversely, if the Slovak subsidiary holds the IP, it must be sold through the bankruptcy estate. Reports indicate the U.S. parent also faces creditor pressure for liquidation, further complicating the destiny of the Prodigy IP.

Swedish venture capital firm Creandum has raised $750 million for its eighth fund, marking the largest fundraise in the firm's history. The raise comes amid a rush by investors to back hot AI startups in increasingly large early-stage rounds.

According to Sifted, Creandum has tripled its presence in San Francisco over the last two years, highlighting its focus on the US AI ecosystem. The fund will provide capital to early-stage AI companies, reflecting the current trend of primary market capital concentrating on top-tier AI projects.

Australian graphic design software company Canva last week released Canva Brush, a new vision-language model designed for designers. Unlike common image generation models, Brush is an editing model intended to understand designer intent and interpret design change requests. According to The New Stack, the model is available to Canva Pro users and can read a design's content and structure, write and execute small pieces of code to make changes, and then check its own results.

In a launch blog, Canva noted that generating high-quality designs is one of the biggest challenges for AI models, but "editing them is even more complex." This is because the model needs to understand what a designer wants to change, how the design is structured, and importantly, what needs to stay exactly as it is. Otherwise, a simple edit request can result in a whole bunch of changes nobody asked for. For example, when handling a prompt like "Make me a 3-page pitch deck for my bakery with Art Nouveau visuals," an agent must make a long chain of decisions, including planning narrative, generating text, selecting imagery, and keeping typography, color, and spacing consistent.

Enterprise AI architect and founder of Besk Tech, Vladimir Beskorovainyi, told The New Stack that Canva is not competing with frontier models but solving a "bounded problem" (one with well-defined specific tasks with clear parameters and limited scope). He believes that on bounded problems, small specialized models have been quietly beating large general ones for years. "The moat here is the artifact, not the weights," Beskorovainyi pointed out, noting that Canva's advantage lies in training on how real designs are actually put together, a training signal harder to match than base processing capability.

Rodion Sorokin, co-founder and CTO of deep-tech venture studio The Capital Index, agreed, saying Canva is "out-niching" the frontier labs. He mentioned that Brush started from an open-weights model and was trained with reinforcement learning against what Canva's internal designers call "good design." Sorokin stated: "Frontier models know the Internet; Canva knows what hundreds of millions of people actually do with a slide deck."

It is worth noting that Canva claims Brush is up to 69% faster than its previous design agent at 46% lower cost per task. However, these speed and cost figures are self-reported by the vendor and compare only with the company's own previous product, not with frontier models in independent benchmarks. Additionally, Canva's other image-generation model, Craft, debuted as the fourth-ranked model on crowdsourced benchmarking platform Design Arena, behind OpenAI's GPT Image series models.

A paper accepted at the top-tier ACM Web Conference 2026, titled "The Algorithmic Self-Portrait: Deconstructing Memory in ChatGPT," utilized data access rights granted under the EU's GDPR to obtain complete conversation logs and all stored memories from 80 real users. The study found that 95.9% of memory entries were written unilaterally by ChatGPT after determining they would be "useful for future conversations," while only 4.1% were explicitly requested by users.

Researchers term this collection of memories an "algorithmic self-portrait," a user profile constructed by algorithms based on self-disclosures in private conversations. Data shows that 28% of memories contain personal data defined by GDPR (such as names), and 7% contain special category data (with health information accounting for 61%). More notably, 52% of memories include inferences about the user's psychological state, such as desires, intentions, emotions, and beliefs. For example, one memory entry reads, "[Name] fears failure, especially regarding major career decisions." This goes beyond factual recording to become AI's judgment of "who you are."

Although OpenAI's official documentation states that the system is trained to "not proactively remember sensitive information unless explicitly requested by the user," in practice, the system allows saving information it deems "useful." This gap between policy promises and actual behavior means users can only indirectly influence AI memory formation through "what they say," rather than directly controlling "what is remembered."

When an aerospace engineer faces a qualified component that is obsolete or has an excessively long lead time, they must spend days cross-referencing vendor datasheets, interface control documents, and qualification records to draft a substitution justification for review. An incorrect substitution on a flight-critical component can lead to mission failure, yet the evidence chain supporting each decision is often scattered across emails and undocumented institutional knowledge, lacking a formal, repeatable package for audit.

The Amazon Web Services machine learning blog describes a multi-agent engineering assistant aimed at compressing this manual cross-referencing process. Its core design rule is singular: engineers make every decision, and the system assembles the evidence supporting that decision. This rule dictates where large language models are used and where deterministic code is employed.

The architecture is defined by three pillars. First, deterministic specification-fit evaluation: Python-written tolerance-rule code compares each candidate component against the reference component and the engineer’s stated requirements item by item. Because identical inputs consistently yield identical pass/flag conclusions, the results can enter formal engineering review. Second, multi-agent orchestration: a lightweight planner routes queries to three specialist agents (Catalog, Evaluation, Knowledge), and a supervisor synthesizes findings only when more than one specialist contributes. Third, human-in-the-loop decision-making: the system drafts justification memos and Integration Impact Assessments for engineering review but does not recommend or approve substitutions. Every evaluation also writes an audit entry aligned with AS9100 sections 8.4 and 8.5.6.

Model tiering represents a pragmatic cost-latency choice within this architecture. Routing is a classification task involving no tool calls, so it does not require a large reasoning model; a Haiku-based planner returns a numbered routing plan in seconds. Specialist agents use Sonnet to execute tools. The Opus supervisor runs only when two or more specialists contribute, because synthesizing a single specialist’s already complete answer adds approximately 8 seconds of latency without quality gain.

The deterministic scoring engine is the foundation of the system’s auditability. Specification-fit evaluation runs as standard Python code against hand-written classification requirement patterns. These patterns define critical specifications and tolerance rules for each component category. The score_substitution_fit function compares items, counts passes and flags, and presents lifecycle risk. Because conclusions must be reproducible, this step involves no large language models. Specialist agents present result tables and explain flags but cannot alter the verdict.

Amazon Nova Multimodal Embeddings assumes two distinct roles in this scheme. In multimodal retrieval, datasheets, interface control documents, and performance charts are embedded into the same semantic space; engineers can input text queries or upload chart images, and retrieval returns matching charts with source attribution. In visual similarity scoring, when two candidate components are compared, Nova Multimodal Embeddings calculates the cosine similarity of their photographs, after which Claude analyzes specific observable differences (mounting hole positions, connector locations, housing geometry). This score provides auxiliary evidence for deterministic specification comparisons, but the verdict itself is based on specification matching.

Circle K operates more than 17,200 convenience stores and fuel retail locations across 27 countries, with cloud-native infrastructure supporting point-of-sale systems, loyalty applications, and electric vehicle (EV) charging services. When alerts fire for these digital services, on-call engineers previously spent 45 to 60 minutes toggling between CloudWatch logs, OpenSearch distributed traces, ECS task statuses, and GitHub runbooks to determine what had happened. Cloud DevOps Engineer Marek Klatowicz co-authored a piece with AWS detailing how they used Agentic AI to eliminate this repetitive labor.

The first version of the system was orchestrated by Amazon Bedrock Flows, coordinating three specialized agents working in parallel and sequence: the AWSResourceExplorerAgent checked infrastructure status, the OpenSearchExplorerAgent pulled application logs and distributed traces, and only after both completed did the GitHubExplorerAgent search code repositories to pinpoint the root cause. A final AnalysisPrompt node aggregated the three outputs into a structured summary containing the root cause, evidence, responsible team, and remediation recommendations, which was pushed to Slack. The entire process took 5 to 15 minutes. However, Klatowicz explicitly noted in the lessons learned section that the first version allowed agents to query CloudWatch logs themselves via action groups, which increased latency and often returned data volumes exceeding processing capacity. The truly effective approach was to have Lambda functions pre-query the logs and place relevant error lines directly into the agent's prompt—agents reason faster when reading actual error messages.

As data volume and knowledge scale grew, the Flows architecture hit hard constraints: some complex events required analysis times exceeding what a single Flow node could complete. The team first attempted optimization within the existing design—more aggressive log filtering, shortened context windows, and splitting services into independent Flows—which helped but was insufficient. Upon migrating to AgentCore Runtime, full investigations ran in a single session, supporting sessions up to 8 hours long, allowing each step to allocate time as needed by the data. The three specialized agents were consolidated into a single agent running on AgentCore Runtime, with AgentCore Gateway centrally managing tool access to OpenSearch, Knowledge Bases, and GitHub.

Klatowicz believes the key driver behind this migration was not AgentCore itself, but an earlier architectural decision: the team consistently kept prompts, tool definitions, and analysis steps separate from the Flow structure. This meant switching from Flows to AgentCore required no rewriting of the triage logic, only changing the runtime method. He also emphasized the importance of search order design: GitHub searches must run after AWS resource and log analysis because the agent already knows which error, component, or recent change to look for at that point; searching the repository blindly at the start returns too many results to be usable.

The cost structure of agent applications differs from single-turn question answering. A shopping assistant completing one recommendation may undergo multiple steps—searching, reading results, narrowing scope, verifying constraints, and submitting an answer—in a multi-step loop, with each step constituting an independent billed call to a frontier model. At scale, this per-token billing loop makes deployment unviable. The solution offered by the AWS blog is sequence-level knowledge distillation (SeqKD): have the frontier teacher model generate complete multi-step trajectories, use a benchmark grader to filter for trajectories that actually solve the task, and then perform supervised fine-tuning on the small model using these trajectories. The teacher only needs to output text, requiring no weights, logits, or reinforcement learning infrastructure, so closed-source API models can also serve as teachers.

The most practically valuable finding in this blog is an ablation experiment: given the same 426 training problems and the same filtering pipeline, the only difference was whether the teacher's reasoning text was retained in the training objective. Removing reasoning and keeping only tool calls dropped the average absolute success rate from 0.541 to 0.415, a difference of about 13 percentage points. The authors explain that reasoning is not decorative narration but a causal chain connecting observations to next actions. Agents need to decide what to search for, how to filter, and when to terminate at each step, decisions driven by reasoning. If the student learns only the sequence of tool calls without the underlying reasoning, it achieves only partial transfer—preserving reasoning is the key to bringing the student close to the teacher's level.

However, SeqKD has a clear capability ceiling. The blog notes that what determines whether distillation succeeds is not the parameter gap between student and teacher, but whether the task itself is within the student's capability range. The teacher GLM-5.3 has approximately 40 billion active parameters, while the student Qwen3.5-4B has only 4 billion—a tenfold gap—but the shopping task is simple enough for a 4B model, so the student almost fully recovers the teacher's quality. Conversely, if the task exceeds the student's original capability, distillation will only produce long reasoning chains that are correctly formatted but contain wrong answers; trajectories lengthen while accuracy remains flat. The authors' recommended response is: upon seeing this symptom, stop adding data, either switch to a larger student or reduce task difficulty. Furthermore, SeqKD plateaus at the hardest tail of multi-step loops because the student cannot see the states entered after its own mistakes; breaking through this bottleneck requires on-policy methods, making SeqKD itself merely the first stage.

From a cost structure perspective, the significance of distillation extends beyond simply shrinking the model. Frontier APIs charge per token, requiring payment for every step of every loop, with costs growing linearly with traffic. The distilled 4B student is deployed on self-hosted endpoints, billed by instance hour, where increased traffic incurs no extra charge until scaling out is required. The authors do not provide specific per-task cost figures because those depend on your request patterns, sequence lengths, and concurrency levels, requiring self-measurement. Meanwhile, IFEval shows the student's general instruction-following capability dropped by about 6 to 8 percentage points, which is the expected cost of specialization—if the drop is significantly larger, it indicates the distillation narrowed the model too severely.