Reading

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.

Since 2000, a total of 557 commercial banks in the United States have failed, with only 28 of them closing on any day other than Friday. Friday is the preferred closure day for the Federal Deposit Insurance Corporation (FDIC), and acquirers are typically arranged before the failure occurs. This means that modern bank failures are rarely triggered by market forces such as deposit runs, but instead follow supervisory closure decisions. For comparison, between 1865 and 1935, before the establishment of the federal deposit insurance system, bank failure dates were distributed almost uniformly across the days of the week—the researchers believe that at that time, depositors and shareholders usually determined the timing of the closure themselves.

But there is an easily overlooked feedback loop here: the basis for supervisors to close banks is book capital (the CET1 ratio), and book capital itself is partly produced by supervisory examinations. The researchers found that the probability of a bank revising its previously submitted financial statements (call reports) sharply increases around on-site supervisory examinations; over the five years preceding a bank's failure, its equity-to-assets ratio typically declines by 8 percentage points, nearly 3 of which come from revisions triggered by examination-induced loss recognition. In other words, supervisors force banks to write potential losses into their statements through examinations, and then use those worsened figures to decide on closure. In the CAMELS rating system, the capital sub-component is mechanically related to book capital, while the management sub-component allows for supervisory discretion—for well-capitalized banks, the composite rating is closer to the management rating; for severely undercapitalized banks, the composite rating is almost equivalent to the capital rating.

The researchers further validated the causal direction using a natural experiment during the 2008 financial crisis: under identical financial conditions, banks assigned to 'stricter' examiners recognized more losses and reported lower book capital, thereby increasing their probability of failure. Stricter supervision indeed reduced the FDIC's resolution costs, but also led to a short-term contraction in credit supply. This constitutes a core trade-off: strict supervision can curb speculative 'gambling for resurrection' behavior and prevent zombie banks from misallocating credit, but it may do so at the cost of further tightening supply when credit conditions are already weak; lenient supervision may preserve credit but raise ultimate resolution costs. The deposit insurance system weakens depositor discipline, making this trade-off even more prominent.

The researchers explicitly state that their analysis is entirely empirical and they make no normative judgments regarding the 'optimal degree of regulatory strictness.' The applicable boundaries of the conclusions are also worth noting: the natural experiment sample is limited to the 2008 crisis period, and the definition of 'strict supervisors' relies on examiners who assign worse CAMELS ratings for the same financial condition—a difference that may partially reflect subjective judgment rather than objective standards. For investors, the value of this research lies in providing an actionable analytical framework: when observing an abnormally high frequency of a bank's financial statement revisions, this is not only a signal of deteriorating asset quality, but also possibly a process where supervisors are actively creating the conditions for closure.

In an investment letter released on November 2, 2026, a16z partners Raghu Raghuram, Martin Casado, and Jamie used a quote from former Netscape CEO Jim Barksdale as their starting point: "The main thing is to keep the main thing the main thing." They argue that most current debates in the robotics industry revolve around training data, model architecture, and hardware types, but there is only one main line—getting robots to complete human tasks. Tutor Intelligence is a company designed around this main line, which is also the core reason they led its Series B round.

Most robotics companies follow a path of building a general model first, then finding application scenarios. Partners point out that this requires massive upfront data collection and capital investment just to build even a basic model. Tutor Intelligence took the opposite approach: identifying real needs and specific scenarios first, deploying hardware into production environments before having a 'robot brain,' and having human teleoperators (whom they call 'tutors') remotely control robots to complete work. Partners stated that these operators were so skilled that customers derived real value from them and could not distinguish between work done by employees and work done by tutors.

Every real-world deployment generates data. Tutor reverse-builds models from this deployment data, enabling robots to gradually gain autonomous capabilities. Partners note that every robot put to work makes the next one smarter, cheaper, and easier to deploy. The data for model training is no longer generic samples pre-collected in laboratories, but operational data truly needed in customers' actual production environments.

In terms of business model, Tutor charges by the hour rather than following the traditional robotics industry practice of multi-million-dollar upfront procurement and implementation fees. Partners believe this shifts capital expenditure to operating expense, unlocking the market for small and medium-sized manufacturing enterprises—where much of American manufacturing exists—and also serving some of the largest logistics companies. According to the investment letter, Tutor's robots are already operating in some of the largest manufacturing and logistics networks in the United States, with plans to have contracts covering over 1,000 robots by the end of 2027; in the past 12 months, nearly one-third of customers who received Tutor robots were existing clients installing additional robots; deployment implementation time is as low as 8 hours.

Tutor currently has two robots: Cassie, which has a single arm and handles case picking and palletizing/de-palletizing workflows; and Sonny, which has two arms and greater dexterity, serving item-by-item picking in e-commerce fulfillment. Both robots share the same model, which is part of Tutor's vertical integration across models, hardware, and data collection. Partners believe this full-stack capability enables Tutor to evolve toward more generally capable robots.

It should be noted that all the above operational data—deployment time, customer repurchase ratio, target number of contracted robots, and claims that operators are indistinguishable from humans—are sourced from self-reports by Tutor Intelligence and restatements by a16z as the lead investor, without independent verification. The investment letter itself includes a disclaimer stating that a16z has not independently verified the information therein. The core of the partners' argument lies in the logic of path selection: whether deploying first and modeling later can continuously produce superior data quality and commercial speed compared to modeling first and deploying later still needs to be tested in larger-scale deployments.

On November 2, 2026, in an interview on The Verge Decoder podcast, Matt Strauss, chairman of the NBCUniversal Media Group, rarely publicly disclosed the specific methods the company uses to evaluate content rights investments. His core formula is not complex: Customer Lifetime Value equals Average Revenue Per User (ARPU) multiplied by the number of months a user remains retained before churning. However, Strauss emphasized that the real difficulty lies in embedding this formula into a portfolio framework—treating every piece of content as a distinct asset within an investment portfolio, separately evaluating whether it drives acquisition, engagement, frequency, or retention, without requiring every individual title to be independently profitable.

He compared this approach to a mutual fund: "Some content may be loss leaders, but it adds value elsewhere in the portfolio." This was precisely the decision logic NBCU applied when its WWE contract came up for renewal. Six years ago when Peacock launched, the team led by Strauss chose to bring in WWE for three reasons: monthly premium live events establish the habit of opening the app; the massive library provides content depth; and the addressable market for WWE was calculated to have sufficient penetration potential. The key judgment was whether the lifetime value of subscribers acquired through WWE could cover the cost of the rights.

When renewal fees increased significantly, NBCU did not feel forced to follow suit. Strauss said they calculated a different question: if we do not renew, how many subscribers can we retain? The answer was the majority—because WWE viewers had already been habituated to other content; they enjoy the Fast & Furious franchise, Yellowstone, and other programs on the platform. Consequently, NBCU abandoned the high-cost renewal, retaining only a much smaller package of WWE programming, and reallocated the saved funds to other parts of the portfolio.

This case reveals Strauss's overall thinking on content expenditure: it is not about "is this title worth the price," but rather "what role does this title play in the portfolio, and where will users go if it is removed." He admitted that the fact that WWE fans like Fast & Furious was initially an assumption that needed testing and verification—"We have assumptions, but we must test these hypotheses."

Strauss put forward a judgment as early as 2019: the streaming market would move from "great unbundling" to "great rebundling," ultimately replicating the economic structure of cable television—dual revenue streams of advertising plus subscription, broad content for all household members, and bundled pricing. He said at the time that Peacock was not built to win the streaming war, but to build growth after the war. By 2026, the market reality he described closely matched this prediction: almost all streaming services have launched ad-supported tiers, 80 percent of Peacock's subscribers are on the ad tier, and the proportion of advertising revenue coming from streaming and sports has risen from about 30 percent six years ago to nearly 60 percent.

As Anthropic prepares for its initial public offering (IPO), one of the most attention-grabbing figures is not revenue, but "$40 trillion." In an analysis published on November 2, 2026, Kim Song-hyun, editor at The Core (South Korea), noted that in the S-1 registration statement filed with the U.S. Securities and Exchange Commission (SEC), Anthropic stated that the economic value of work performed by global knowledge workers exceeded $40 trillion last year. While Anthropic did not explicitly label this figure as a "Total Addressable Market" (TAM), the article argues that it effectively transforms the value of productivity gains AI could deliver into a business opportunity within investment narratives.

What truly warrants close examination here is the shift in market definition. Historically, companies have typically defined their markets based on what they sell—for example, estimating the spending scale around current products and services. Anthropic’s approach differs: it takes human work that AI can replace or augment as its starting point. The economic value of knowledge labor—including development, research, accounting, legal services, and office work—is included in the potential market. AI companies are no longer just selling model API calls or software seats, but rather execution capabilities that enter enterprise workflows.

The Core compares this shift to previous public listings. Companies like Snowflake estimated markets largely based on existing product spending; Airbnb and Uber incorporated the economic activity connected by their services into their market definitions; and in discussions related to its listing, SpaceX presented a market opportunity where AI accounted for a significant portion. Anthropic’s $40 trillion figure appears more aggressive because it uses the "value of human labor" itself as the market boundary. The larger the boundary, the grander the story may seem, but it also risks obscuring a key distinction: the creation of economic value does not equal a company’s ability to capture revenue.

The article illustrates this distinction with a simple example: if AI helps a company save 10 billion South Korean won in costs annually, the price the company pays for AI services may be far lower than 10 billion won. A significant portion of those saved costs might become profits for the adopting company or be passed back to customers. Therefore, even if AI has a substantial cumulative impact on the global economy, how much revenue Anthropic captures depends on pricing, competition, customer budgets, and switching costs. The Core also cites IDC forecasts suggesting that investments in AI solutions and services could generate a cumulative impact of $22.5 trillion on the global economy by 2031, but this remains a macroeconomic productivity metric, not Anthropic’s order book.

From the operational numbers disclosed in the S-1 analysis, Anthropic’s growth is indeed rapid. The article states that Anthropic’s 2025 revenue was $4.585 billion, while Q1 2026 revenue reached $4.731 billion, already surpassing the full-year 2025 total. Q2 2026 revenue was $11.555 billion, bringing the first-half total to $16.286 billion—approximately 14 times the same period last year. Meanwhile, the company still reported an operating loss of approximately $693 million in the first half, but the loss rate dropped significantly from about 184% in the prior year’s first half to roughly 4%. In Q2, GAAP operating profit was approximately $974 million, with an operating margin of about 8.4%.