Reading

Ethan Mollick recently produced a set of data charts for his new book Co-Existence. The tool he used was Opus 5.5 inside Claude Projects—an 'orchestrator' agent that, upon receiving instructions, splits tasks on its own, spins off sub-agents, assigns work, and relays results back to him. He gave a couple of sentences of suggestions, and the AI completed what would have taken him hours: all the research, all the graphing, all the layouts, all the images, all the citations. He even told the AI to 'make it better'—a vague reaction that would drive a human team mad—and the AI complied. More surprisingly, it caught a math error he had made.

This experience itself became the starting point for his argument. In Co-Intelligence, published in early 2024, Mollick described two modes of human-AI collaboration: 'centaur' work (divided between AI and humans, each doing their own parts) and 'cyborg' work (blended between AI and humans). Both keep you in the loop. But now that we have real agents, he added a third: the 'conjurer.' Conjurers summon agents when needed, assign them work, and evaluate what comes back. The way he made these charts was conjurer-style: summon agents, give some general guidance, and say 'make it better' repeatedly.

Why has this shift happened? Mollick uses the METR Long Horizon Benchmark to illustrate. The benchmark measures the amount of human work an AI can do with at least a 50% chance of success from a single prompt. From the release of GPT-4 to the end of the METR measure, that number doubled eight times. In 2024, an AI could do a four-minute task; by April 2026, when he turned in his draft, an AI could do 17 hours of human work. Then the benchmark itself ran out—METR said they could no longer accurately measure the amount of work an AI could do using their current methods. Other studies suggest that number might be 17 weeks of work, or more.

Costs have dropped in parallel. The Google Proof Question and Answer Exam he mentions—giving AI problems that regularly stump graduate students—used to cost 30 cents in AI usage to get a single correct answer; now it costs four hundredths of a cent. The model landscape has shifted too: GPT-4's period of undisputed dominance lasted roughly a year, after which twenty different models have held the lead at one time or another, all from Anthropic, Google, or OpenAI. A total of 180 AI models are now better than GPT-4 was at its release, many of them open-weights models that anyone is free to use or modify. Nineteen can run on a modern smartphone.

But Mollick emphasizes that the frontier of AI ability is still jagged. Just as it was 30 months ago, some areas have stayed stubbornly weak—AI is still quite a bad long-form writer—while others, like solving math problems, went from terrible to superhuman in the same period. And even where AI is excellent, small gaps create large needs for humans—at least for now.

The risk of the conjurer mode lies precisely here. It is the most powerful way to use AI, but we lose control over what, exactly, the AI is doing to get us those results. You do not know what the agent did in the middle, what it skipped, or where it might have gone wrong. Mollick himself says he told the AI 'make it better' many times—a vague reaction that would drive a human team mad, but the AI complied.

His new book Co-Existence does not try to keep up with the news or the latest model releases. Rather, it focuses on trends: AI is improving exponentially but unevenly, and the way we use it is shifting from chatbots to agents that do real work. The book's core question is what that means for the humans who have to co-exist with these systems: where we still have an advantage. His answer comes down to four things—deep knowledge, wide knowledge, taste, and agency. If you understand the trends, the present and near future become much less surprising.

ElevenLabs builds the voice layer of AI—models that turn text into speech that sounds human. Most people have already used it when talking with customer service, without realizing it. Klarna runs first-line phone support for 35 million U.S. customers on it. So do Deutsche Telekom, Cisco, Adobe, and a growing list of governments. The company says it is pacing at $600 million in annual recurring revenue and is now reportedly valued at $22 billion by its backers, despite being just four years old.

ElevenLabs co-founder and CEO Mati Staniszewski was asked about gross margins during an interview with TechCrunch Editor in Chief Connie Loizos at the Nrth conference in Toronto in September 2026. His answer was direct: he does not mind margins getting squeezed even further if it means expanding the company's market share. He said that given the research element, the company is able to fine-tune and constrain models in extremely smart ways to optimize costs, but if savings can be passed on to the customer, they do that. In his view, the biggest thing right now is still proving value and being there with the customer, and once value gets created over the next five years, everyone benefits together.

This statement only has concrete meaning within ElevenLabs' competitive landscape. Staniszewski acknowledged in the interview that the company is increasingly bumping into its own customers—Decagon, for example, is a conversational AI platform that trained its voice product on ElevenLabs and now competes with it by running queries through its own models. Staniszewski believes the lines between model companies, platform companies, and application companies are becoming blurrier. He cited Anthropic as an example: what was a model company is definitely a platform and increasingly a wide set of applications.

At the same time, Staniszewski revised his judgment on audio model commoditization. He had said at TechCrunch Disrupt the previous year that audio models would be commoditized within a couple of years. Now he acknowledged that the quality delta achievable at the model level is still significant, but pushed the timeline to three to five years. In other words, he believes the technical moat has a window, but not a permanent one.

Putting these together, Staniszewski's strategic logic is: since the model quality gap will narrow within three to five years, and customers can become competitors at any time, trading margin for customer lock-in and integration depth is rational at this stage. ElevenLabs' other moat lies not in the model itself but in data annotation—Staniszewski said the big part of training has not been so much the volume of data but annotating it. The company has thousands of people internally on a contracting basis helping annotate not only what was said, but when people were speaking, how they said things, and what emotions were used. They even brought in voice coaches to detect accents accurately. This operationally intensive data asset is not something open-source models can directly replicate.

However, the boundaries must be clear: Staniszewski declined to disclose specific gross-margin figures. The $600 million ARR is self-reported by the company. The $22 billion valuation is reportedly from a Bloomberg article in July 2026 about tender-offer talks. His statement about not minding margin compression is an expression of strategic intent, not a realized financial result. Whether the promised mutual benefit materializes over the next five years depends on whether the company can recover costs through pricing power after expanding market share, or whether margins get compressed to the point of being unable to cover research and annotation investment.

GPT-OSS-20B is an open-weight reasoning model with a stubborn habit: regardless of which language the user asks in, its chain-of-thought almost always generates in English. The AWS machine learning team found that on a 50-example multilingual reasoning test, the base model used the correct target language only 22 percent of the time, and that 22 percent came almost entirely from the English samples—French, Spanish, and Italian scored zero, with German at just 10 percent.

The team then ran supervised fine-tuning (SFT) for 1,000 steps using a LoRA adapter (r=8, α=16) across four ml.g6e.12xlarge instances. The result was direct: reasoning-language accuracy jumped from 22 percent to 100 percent. But answer-language reached only 62 percent—the model learned to think in French yet often switched back to English for the final answer.

This is precisely where GRPO (Group Relative Policy Optimization) applies. GRPO works by sampling a group of responses per prompt and pushing the policy toward the above-average ones. If the target behavior has zero probability in the base distribution, there is almost nothing for GRPO to reinforce. After SFT, the model could already reason in the correct language 100 percent of the time and occasionally produce the answer in the right language as well, giving GRPO a signal to reinforce.

The core asset of a quality-management platform is institutional memory: a similar defect investigated last year, a Corrective and Preventive Action that fixed the root cause, an audit finding that flagged the same gap. That knowledge sits scattered across billions of documents, spanning languages and formats, and no human can search by semantic similarity, so every investigation starts from scratch. The Octave Reliance team saw unlocking this history as its single largest opportunity, and also saw the obstacle: complaint records, audit findings, and corrective actions carry the platform's strictest isolation requirements, and those were exactly the records the AI would need to read. In a platform serving over 14,000 customers across medical devices, pharmaceuticals, and food manufacturing, a single inadvertent cross-tenant disclosure would carry regulatory consequences far outweighing any AI benefit.

Faced with that contradiction, the team made an unusual call: rather than shipping one AI feature, it spent its first cycle building a foundation every future feature would sit on. The reasoning was straightforward—Reliance has over 40 applications, each with different document types and workflows, and building AI for one in a silo meant rebuilding the infrastructure for the next. So they inverted the plan: build the horizontal platform first, then cut vertical feature slices into it. Complaint summarization and triage went first, to prove the architecture before the rest followed.

Four tensions shaped the design, and none could be traded away: making data available without risking inadvertent disclosure, executing agents in isolation, reading content that cannot be trusted, and embedding 2 billion documents without a quarter-long wait. When all four had to hold simultaneously, the option space narrowed considerably.

The first key decision was where to enforce tenant separation. Application code would have worked, but it would have meant every future feature team getting it right on every code path, where one missed check risks inadvertent cross-tenant disclosure. The team pushed isolation into the runtime instead: each user session runs in a dedicated microVM with its own CPU, memory, and filesystem, stopped and sanitized when the session ends. That is hardware-level separation designed to hold regardless of how the agent behaves. AgentCore Gateway authorizes at the tool, operation, and parameter level, and automated reasoning provides formal verification against the defined rules.

The second key decision was how to represent the data. The team evaluated embedding models against its own corpus rather than public leaderboards. The deciding questions were specific to this data: what a photo of a cracked housing has to do with the paragraph describing it, and a German complaint with an English CAPA. Cohere Embed v4's multimodal capability places photos and text in one vector space, and native multilingual support means documents in different languages within one tenant's corpus need no translation preprocessing. But the finding that only a proprietary corpus could surface was the dimension choice: Amazon OpenSearch Service costs scale linearly with dimension count, and Embed v4 supports configurable output dimensions of 256 to 1,536. Testing against representative production queries showed 512 dimensions holding target accuracy while cutting storage roughly 66 percent against the 1,536 default. This trade-off would not appear on any public benchmark, because it depends on the actual document mix of defective-product photos, scanned inspection forms, annotated diagrams, and text.

On results, the Complaint Triage Assistant has been running in production for three months. Octave measures 4 to 6 hours per investigation as the work the pipeline absorbed: extraction across formats, the hunt for precedent, and the reading it takes to tell a genuinely similar defect from a coincidence of wording. The foundation cut deployment time per use case by 50 percent—by Octave's own accounting—and semantic search landed in Q2, two quarters ahead of the original Q4 target.

Boundaries worth preserving: all figures come from Octave's internal testing and production self-reporting, published on the AWS official blog. The latency and cost overhead of the microVM approach is not discussed in the source. The team also drew a line it chose not to cross: the assistant surfaces context and precedent, and the quality handler still makes the call—in an industry where that call carries regulatory weight, that boundary shaped what got automated and what did not. This is one company's architecture experience in a specific regulatory context, not a proof of optimality for runtime isolation across all multi-tenant AI scenarios.

Figma announced 20 new MCP partners at its first Config conference held in India, including Amp, ClickUp, Linear, Notion, Mistral, and others. At a media briefing, co-founder Dylan Field told reporters that he does not think MCP is a perfect protocol.

Field's core argument points to a structural asymmetry: MCP was created by Anthropic, yet Anthropic itself does not expose its cloud resources through MCP. Field wrote on X that he found it interesting that the creators of the MCP standard do not expose artifacts or other cloud resources created via MCP. He argued that the debate is framed as open versus closed, everyone always prefers open, and meanwhile issues of asymmetric context sharing are ignored.

The mechanism behind this argument is as follows: MCP standardizes how AI agents connect to external data sources and tools. When application-layer companies open their data according to this standard, the protocol creator—typically a frontier lab with the most powerful models—gains a structured access path without offering reciprocal exposure. Field further stated at the media briefing that MCP is advantaged more to frontier labs.

Figma's actual strategy is to both embrace and critique the standard. The company is expanding MCP support, but Field said the reason Figma does not want to be fully open is that the more data the company exposes through MCP, the higher the chance of abuse. He acknowledged that most users do not intend to use MCP or Figma data maliciously, but the company still has to think about those potential risks. Field later added that Figma plans to support Pi, the open-source tool that one of MCP's co-creators had complained about.

The value of this argument lies not in whether it ultimately proves correct, but in that it raises a concrete design question: when a protocol is created by the most powerful participant in an ecosystem, how do smaller companies adopting that protocol avoid becoming one-sided data suppliers? For developers evaluating AI agent integration options, the asymmetric context sharing Field identifies is a practical constraint to consider, not merely an ideological open-versus-closed debate.

The US State Department announced on October 20, 2026, that it has added Shanghai Zhisuan Technology Co., MiraclePlus, and three affiliated companies to its sanctions list. According to Bloomberg, the action stems from allegations that these entities were linked to the case involving former Google software engineer Ding Linwei, who was convicted this year of stealing artificial intelligence trade secrets and sentenced to one year in prison.

Zhisuan, founded by Ding, develops a software platform for training AI models. The State Department stated that Ding applied on behalf of Zhisuan to participate in MiraclePlus's startup incubation program, emphasizing his knowledge of Google's technologies. MiraclePlus is the successor to Y Combinator's China program, established by Lu Qi, a former Microsoft AI scientist. The sanctions specifically target MiraclePlus for providing seed financing to Zhisuan.

Under the restrictive measures, assets held by these entities in the United States may be frozen, and their access to banking services will be restricted. This move signals that US regulators are extending enforcement tools to venture capital firms involved in funding startups with questionable intellectual property origins, rather than targeting only the direct thieves or end-users.

According to Bloomberg, four co-founders of Chinese artificial intelligence company Z.AI Co.—Li Juanzi, Tang Jie, Xu Bin, and Zhang Peng—participated in a research project between 2018 and 2021. The project was jointly funded by the National Natural Science Foundation of China and the China General Technology Research Institute (CGTRI), with an approved budget of 2.94 million yuan.

Britain’s Security Service (MI5) previously issued an alert stating that CGTRI has “very strong ties” to China’s Ministry of State Security (MSS) and funds academic work intended to strengthen the MSS’s espionage capabilities. MI5 advised UK academic institutions to immediately review any collaborations with CGTRI.

Citing a 78-page project report, Bloomberg said the research envisioned AI systems capable of mining social media posts to identify emerging events and reveal links among people. The report emphasized that this description concerned proposed capabilities, not evidence that such tools were deployed.

China’s Embassy in Britain disputed MI5’s allegations against CGTRI. Z.AI and the four founders did not respond to Bloomberg’s requests for comment. Filings by Z.AI with the Hong Kong Stock Exchange confirm the four individuals as co-founders, with Zhang serving as chief executive.

According to the Wall Street Journal, the US Department of Homeland Security has issued an order banning federal civilian agencies from using artificial intelligence technology from the Chinese lab DeepSeek. The directive requires these agencies to remove existing DeepSeek products from their IT systems and prohibits the purchase of any new ones.

The ban extends beyond DeepSeek's own products to include offerings from affiliated companies and third-party services built with DeepSeek implementations. However, the notice did not set a specific compliance deadline or detail the particular security concerns that led to the decision. It remains unclear what "built with" actually means in practice, leaving questions about whether AI gateways like OpenRouter fall under the restriction.

The order is based on a 2018 US law covering supply chain security in federal procurement. This move follows earlier discussions within the US government about taking steps against powerful Chinese open-weight models, such as Kimi K3 from Moonshot AI. While major tech firms like Nvidia, Microsoft, Google, and Meta previously objected to broad restrictions, OpenAI and Anthropic had lobbied for tighter controls, marking a shift towards concrete regulatory action against specific foreign AI providers.

Can activation steering compel an agent to take actions, rather than merely talk about a topic? This is the extension question that LessWrong author Ephraiem Sarabamoun posed after prior steering experiments like Golden Gate Claude had already shown the technique could control text output. He ran Qwen3.8-27B as an OpenClaw agent in a sandboxed environment, giving it access to tools for light control, ending the chat, and other basic actions. Experiment 1: After building color steering vectors and injecting them, the agent consistently turned the lights to the specified color; increasing the steering dosage made the agent tend to output the steered color instead of the color it was instructed to produce. Experiment 2: A similar approach built a steering vector for the end-chat tool; once injected, the agent triggered the end-chat button regardless of what action it was asked to perform. The author considers this direction potentially useful for AI safety and plans to continue exploring with more powerful agents.

Stijn Van Nieuwerburgh, a professor of finance and real estate at Columbia Business School, argues in a paper prepared for a Brookings Institution conference that the economic output consumed by US artificial intelligence infrastructure construction may exceed that of historical mega-projects like electricity and railroads. According to Reuters, the study projects that by 2032, approximately 3.6% of US GDP will be invested annually in this sector, with a cumulative scale exceeding $10 trillion.

This ratio is significantly higher than the peak of US railroad construction in the late 19th century, which accounted for about 2.2% of GDP annually. In comparison, the Interstate Highway System started in the 1950s and the telecommunications expansion in the mid-1990s both saw annual investments slightly above 1% of GDP. Van Nieuwerburgh points out that previous expansions in rail and telecom triggered asset bubbles and subsequent crashes. The current AI infrastructure boom not only has unprecedented scale but also increasingly complex financing structures involving banks, private credit institutions, and special purpose vehicles (SPVs). This opacity evokes comparisons to the pre-subprime crisis era.

To test the viability of these investments, Van Nieuwerburgh calculated the required revenue gap. He estimates that by 2032, the AI industry needs to generate approximately $3.7 trillion in total annual revenue to achieve currently expected investment returns. However, the combined annual revenue of leading companies like OpenAI and Anthropic is currently estimated at only about $100 billion. This implies that related revenues must grow at a rate of roughly 80% per year to bridge this massive gap. If market demand remains uncertain, technology changes too rapidly, or execution bottlenecks occur, high leverage rates could lead to "substantial downside risk" upon a shift in market expectations.

However, the professor emphasizes in the paper that this does not mean a financial crisis is imminent. He suggests that strong growth in AI applications, high infrastructure utilization rates, and continuous improvements in model capabilities could support the anticipated infrastructure investments and generate stable cash flows. The paper was scheduled for formal submission on Friday, with its core message serving as a reminder to investors and policymakers to monitor the tension between current capital expenditures and future revenue potential.

When AI agents access databases, APIs, and sensitive business operations through the Model Context Protocol (MCP), a fundamental question arises: how to prove which agent made a request, prevent forgery between agents, and confirm that each agent acts only on behalf of an authorized user. MCP has no built-in authentication standard and is transport-agnostic by design. Without proper identity mechanisms, any component could impersonate an agent, replay intercepted credentials, or invoke tools without authorization.

The AWS and Keyfactor engineering teams jointly published a reference architecture that addresses this gap through four complementary security mechanisms. First, signed request envelopes provide cryptographic proof-of-possession on every agent call. Second, user-signed delegation assertions authorize agents without sharing credentials. Third, RFC 8693 audience-scoped tokens limit the scope of exposure from any single token disclosure to one service boundary. Fourth, real-time OCSP revocation checking propagates credential changes immediately across all agents.

The revocation cascade is the architecture's key security property. When a user logs out, the system first revokes the user certificate in EJBCA. Because delegation verification checks the user's certificate through OCSP in real time, this step immediately invalidates all outstanding delegations. The system then revokes all three agent certificates, invalidating every signed envelope, deletes all secrets from AWS Secrets Manager with no recovery window, and terminates the Keycloak session. The critical insight is that revoking a single user certificate at the PKI layer promptly revokes all downstream agent authority without needing to individually revoke tokens, delete sessions, or contact each agent. OCSP provides the propagation mechanism, and because every request checks OCSP, the effect is immediate.

On private key isolation, the user's private key is used only at login to sign the delegation assertion and is never shared with agents. Each agent's private key is used only by that agent to sign envelopes and is never shared between agents. EJBCA never sees or stores private keys, operating purely as a certificate authority. If Secrets Manager is subject to unauthorized access, the PKCS12 password stored separately adds an additional layer of protection.

The architecture is framework-agnostic. While the demonstration uses the AWS Strands Agents SDK and FastMCP, the same PKI patterns apply to LangChain, CrewAI, AutoGen, or any other multi-agent framework that uses tool protocols. As AI agents proliferate and gain access to increasingly sensitive operations, agent identity infrastructure moves from optional to essential.