• Altman said AI models are improving rapidly at finding real-world vulnerabilities and that broad scans of internet code may uncover more serious flaws than expected; he argued that durable defense may require re-architecting software, not only patching bugs. Cheeky Pint He defended iterative deployment: put advances into people’s hands only after making them safe, rather than building up a capability overhang. Cheeky Pint
• He favors many slightly superhuman AIs over a single vastly more powerful one, arguing that interacting agent teams may have useful scaling and alignment properties and that mutual checks, pluralism, and resilience matter for safety. Cheeky Pint
• On AI-assisted AI research, Altman said models can already be asked to find a 5% compute-efficiency gain, but researchers do not yet know how to elicit a Transformer-scale architectural advance; he identified high-level planning and conceptual reasoning as gaps and said OpenAI’s approach remains experimental. Cheeky Pint
• Altman said OpenAI has made a chip with high intelligence performance per watt, is learning to build its own data centers, and plans to direct its robotics program toward building them; he framed AI-directed robots, fabs, and energy infrastructure as a prospective route to more compute, while noting permitting and components as bottlenecks. Cheeky Pint

• Hassabis said AlphaFold predictions had reached 250 million protein structures and were used by more than 4 million researchers; Isomorphic Labs was at a preclinical stage, with the longer-term aim of reducing drug-discovery timelines from many years—perhaps a decade—to months or weeks. He cautioned that protein structure is only one part of drug discovery. Casina Pio IV
• He described transformer-based deep reinforcement-learning systems as using learned models to guide search, and argued that verifiable fields such as mathematics and coding can support answer-checking and synthetic-data feedback that compounds across training generations. He expects general models such as Gemini to orchestrate specialist tools such as AlphaFold and AlphaGenome. Casina Pio IV
• Hassabis said current AI is not AGI, citing uneven performance and a lack of genuine invention. His proposed “Einstein test” would ask a model trained only on knowledge available through 1901 to conjecture special relativity; he said today’s systems cannot do this, but expected the capability to emerge within two to four years. Casina Pio IV
• He highlighted experimental validation as a bottleneck for AI in natural sciences: one system proposed 200,000 new materials, but only a tiny fraction had been lab-tested; automated labs and simulations could help narrow the search. Casina Pio IV
• Amid commercial and geopolitical pressures, Hassabis called for rigorous scientific practice rather than “move fast and break things,” a standards body to test frontier AI before public deployment, and international cooperation on a shared global framework. Casina Pio IV

• Altman said model progress had accelerated over the prior six months; GPT-6 marked a step forward in discovering knowledge and tackling difficult tasks, and had OpenAI’s strongest adoption of any new model to date. CNBC Television
• Altman said broad enterprise AI rollouts are bringing software creation to employees who previously did not see themselves as developers, enabling them to make apps and other work artifacts. CNBC Television
• He urged companies to prioritize understanding and mitigating AI-related security threats. Altman strongly agreed that AI is exposing longstanding bugs rather than creating them, and said addressing vulnerabilities may require automated fixes and services that continuously monitor traffic instead of patching every bug individually. CNBC Television
• Altman said AI is helping with scientific discovery and chip design; he said OpenAI’s Jalapeno chip was being deployed in OCI, called it the world’s best inference AI chip, and said AI accelerated its development. He predicted AI-driven innovation could yield higher-quality, lower-cost services. CNBC Television

• Sam Altman said AI progress had accelerated, describing GPT-6 as a step forward in knowledge discovery, difficult tasks, and completing substantial work. He called its adoption OpenAI’s strongest for any new model, citing use across coding, enterprise software, security, science, and other industries. CNBC Television
• Altman said broad enterprise rollouts quickly lead employees who do not think of themselves as programmers to build apps and create work artifacts for their roles. CNBC Television
• He urged companies to prioritize understanding and mitigating AI-related security threats, even over near-term productivity gains from Codex. Altman endorsed Oracle co-CEO Clay Magouyrk’s point that AI is uncovering existing vulnerabilities rather than creating bugs; he warned that patching flaws one by one may not suffice and said new approaches, such as continuous traffic monitoring, may be needed. CNBC Television
• Altman said AI is helping advance science and technology, citing AI-assisted chip design. He said a chip called Jalapeno was being deployed in OCI, called it the world’s best AI chip for inference, and said AI had accelerated its development. CNBC Television

• LeCun said AMI Labs began operations in March, had raised about $1 billion in seed funding and had around 65 employees; he described its focus as physical-world AI built around world models. Evident Insights
• He argued that superhuman performance on symbolic tasks does not amount to human- or animal-like intelligence: current AI approaches struggle more with messy, continuous real-world tasks, a gap he illustrated with the absence of domestic robots and Level-5 self-driving cars. Evident Insights
• LeCun defines a world model as predicting how the world changes after an action, so a system can plan toward a goal. He said a paper on hierarchical world models had been published the prior week; his long-term proof point is a humanoid able to do tasks a 10-year-old can manage without extensive demonstrations. Evident Insights
• He advocated learning abstract representations and predicting what is predictable rather than every detail; for financial data, which he described as mostly noise, he said systems should account for uncertainty rather than predict one specific outcome. Evident Insights
• LeCun said much of the world-model and joint-embedding research he sees comes from Asia, especially China, and characterized China as catching up faster than Silicon Valley, which he sees as too focused on LLMs. Evident Insights

• LeCun presents JEPA as an alternative to generative AI: it predicts in abstract representation space rather than trying to predict pixels, and he says JEPA methods work for self-supervised vision and video and can support world models that predict the consequences of actions for planning. TBS CROSS DIG with Bloomberg He characterized LLM-style VLA/VLM approaches to robotics as not working and said many practitioners were moving toward world models, which he described as a major research focus. TBS CROSS DIG with Bloomberg
• He argues that LLMs cannot reach human-level intelligence without a major architectural change: they lack physical-world understanding and explicit world models for anticipating the consequences of actions. He describes LLMs as useful for distilling knowledge expressed in text, but says much human knowledge—including common sense—is nonverbal. TBS CROSS DIG with Bloomberg
• LeCun said he left Meta as his world-model techniques began working, but their main applications were in industrial control, manufacturing and robotics, while Meta had de-emphasized robotics and refocused on LLMs around spring 2025; he said he set up a company because this research requires expensive people and GPUs. TBS CROSS DIG with Bloomberg He identified industrial uses including predicting system states, anomaly detection, predictive maintenance and optimal control, with robotics and autonomous driving as longer-term applications. TBS CROSS DIG with Bloomberg
• On safety, LeCun called AI-extinction probability estimates unsupported and rejected the assumption that greater intelligence necessarily leads to domination. TBS CROSS DIG with Bloomberg He also called LLMs intrinsically unsafe but not very dangerous because they are not very smart, while arguing that objective-driven systems with world models and guardrail objectives could be made controllable and safe. TBS CROSS DIG with Bloomberg
• He supports open, frontier foundation models as a basis for diverse personalized assistants, and warns that reliance on a few proprietary AI assistants could concentrate power and make people’s information diets—and human culture—more uniform. TBS CROSS DIG with Bloomberg

• On Moonshot AI’s Kimi K 3.1 Pro, Emad said it appeared to distill Anthropic’s Opus 5.5 while extending Moonshot’s Kimi 3; he saw capabilities converging among labs outside perhaps Europe, and highlighted its ability to generate 3D game assets directly in code. Peter H. Diamandis
• Emad described GPT Terra and Sonnet 5 as models being developed for personal-companion use: reliability and lower hallucination rates matter, while the user-facing model could route work to specialized models through a control-plane setup. Peter H. Diamandis
• Emad argued that excluding biological or chemical dangerous knowledge from pretraining need not reduce general capability and may slightly improve it; malicious code can hurt performance and be difficult to train back in, so he favored omitting such material in pretraining and adding needed capabilities later through post-training. Peter H. Diamandis
• On Anthropic’s proposed IPO, Emad argued the company did not need the cash: he cited $130 billion raised, $110 billion in cash and equivalents, and $10–20 billion spent, and said another $100 billion raise would put its cash balance around $200 billion. He said Anthropic should have stayed private to avoid the conflicts of being publicly traded while prioritizing its mission. Peter H. Diamandis
• Reacting to OpenAI’s announced foundational results in mathematics, computer science, and physics, Emad questioned how AI-assisted, formally verifiable work should be credited, criticized the unreleased physics result, and urged using AI to extend research and explore new approaches. Peter H. Diamandis

• Amodei’s AI-safety work at OpenAI treated failures as engineering problems: he was first author of the 2016 paper Concrete Problems in AI Safety and co-wrote an early paper on training AI from human preferences. Hushmere He was also last author on the January 2020 scaling-laws paper. Hushmere
• In February 2026, Anthropic refused to drop Pentagon-work limits against mass surveillance of Americans and fully autonomous weapons. Hushmere A California judge later found federal measures against Anthropic unlawful retaliation for its speech, while a Washington appeals court upheld a separate Pentagon exclusion under a 2018 law; Anthropic sought rehearing. Hushmere
• In an August post, Amodei said the strongest criticism of AI companies, including Anthropic, was that they had not delivered on promises to benefit the world. He called “AI will cure cancer” a cliché and said the meaningful achievement would be actually curing cancer. Hushmere

• Joelle Pineau rejected the idea that frontier AI can be built only in the US or China: she said the methods and ingredients are reasonably well understood, with access to capital for compute a main bottleneck and talent another; she cited Canada’s research ecosystem and said Cohere’s recent funding round would help it scale, while stressing that model training requires careful technical choices. MTS
• Cohere aims to stay in the frontier-model range without over-optimizing for shifting index rankings, prioritizing value for money. Pineau said many enterprise customers do not need expensive frontier models, so Cohere’s model-agnostic platform matches models to workload and balances capability, cost, and security; its enterprise deployments can also run on-premises in air-gapped environments. MTS
• Pineau described AI sovereignty as the ability to control access to technology, its cost, and where it runs; she said Canadian-built models running in Canadian data centers are an asset, and noted that Canada’s government is investing in Cohere’s round. MTS
• On allied model development, Pineau said Cohere had model generations in pretraining, fine-tuning, and evaluation, and identified a Germany-based partner for core-model work and its Prague-based cybersecurity team for cyber work and evaluation protocols. MTS
• Pineau said Cohere is testing a cybersecurity partner’s vulnerability-finding harness with Cohere models, including smaller models she said can be effective at detecting vulnerabilities. She urged companies, governments, and nonprofits to use AI to strengthen their own defenses as attacks grow in volume and complexity. MTS

• LeCun argued that current AI remains weak at handling the physical world despite strong performance on discrete-symbol tasks; he said text-only training will not produce human-level intelligence. Dartmouth
• His proposed objective-driven architecture encodes observations into a state, predicts future states under candidate actions, and uses optimization to choose actions, with guardrail objectives intended to keep planned trajectories safe. Dartmouth JEPA predicts abstract representations rather than reconstructing every detail, aiming to discard unpredictable information. Dartmouth He also described a three-level model that improved planning success over non-hierarchical planning at the same compute in a synthetic ant-like-agent task. Dartmouth
• LeCun said he had created AMI Labs to develop physical-world AI using world models and planning—not language—with applications including biomedical science, healthcare, manufacturing, and physical science. Dartmouth He predicted another AI revolution in physical-world systems, with domestic robots as a visible sign, but said it would require conceptual breakthroughs and would not happen immediately. Dartmouth

Google DeepMind’s AlphaEvolve found a method for multiplying 4×4 tables of complex numbers using 48 multiplications, one fewer than Strassen’s 49; Demis Hassabis described the approach as combining LLM-generated candidate solutions with evolutionary computing to explore for novel solutions. Code Bear

• Yann LeCun argues that current LLM and generative approaches handle symbols more readily than high-dimensional, continuous, noisy sensor data and do not anticipate the consequences of physical actions. He says human-level intelligence requires quickly adapting to unfamiliar tasks with little or no prior training, cannot come from text alone, and advises students to focus on physical AI rather than LLMs. ビジネス＋ITチャンネル
• LeCun proposes JEPA, which predicts in a learned abstract representation rather than trying to reconstruct unpredictable video pixels; he describes multi-level versions that predict over different timescales to support planning, and suggests learned world models for complex systems such as jet engines, chemical or power plants, patients, and cells. ビジネス＋ITチャンネル
• His proposed world-model agent predicts the state resulting from candidate actions and optimizes an action sequence against a task objective. He adds a runtime safety objective and presents this as an intrinsically safe agentic-system blueprint because the system predicts action consequences before carrying them out. ビジネス＋ITチャンネル

• Emad argued that practical access matters more than frontier rankings: most work will involve either capable models for everyday tasks or large swarms of 1,000–10,000 agents for complex work, with little in between. Peter H. Diamandis
• He forecast that search advertising could peak in Q1 as agents handle shopping and SEO, putting Google and Meta at risk if they do not own the AI assistants closest to users; he contrasted Gemini Spark with Dots and Grokbot, which he said had a million users. Peter H. Diamandis
• Emad said generative AI models learn physics and can feed insights back into real-world systems, potentially accelerating understanding of phenomena through large-scale training runs. Peter H. Diamandis
• He argued that human conviction helps drive ambitious action, while current AI systems tend to explore rather than persistently pursue goals; he attributed this to how models are built and said AI could be designed with greater conviction. Peter H. Diamandis
• On his own engineering practice, Emad said he had not looked at a line of code in a month despite building “state-of-the-art things”; he said he did not know how much code he had generated. Peter H. Diamandis

• Joëlle Pineau said she had watched Montreal graduate students leave for the U.S. after her own AI research work at Meta and academic experience at McGill and Mila. She described founding Cohere in Toronto as a major act of conviction and risk, and said its current ambition depends on raising capital, securing compute, and assembling talent while keeping a Canadian base. Evan Solomon
• Pineau said Cohere aims to become a global AI company and needs partners who understand local culture, language, and markets. She cited the roughly 30-person Reliant AI team, based between Montreal and Berlin, joining Cohere earlier in 2026, and described a larger German merger bringing integrated modeling teams and experience deploying AI in industry and government. Evan Solomon
• A Cohere colleague credited Pineau with leading much of the company’s work across AI safety concerns including labor-force impacts, bias and inclusion, and model performance across the languages of the populations served. Evan Solomon

• Arthur Mensch described Mistral as building “sovereign AI” for enterprises through frontier models, agents and business applications, with technology deployable on customer infrastructure or hosted by Mistral. He said supply constraints also led the company to own and operate GPU infrastructure, integrating from server deployment through customer applications. Deirdre Bosa
• Mensch argued that open-source models support enterprise business continuity through provider and infrastructure portability, protect control over intellectual property, and reduce costs. He said Mistral’s cyber capabilities were state of the art against Chinese models and argued Western open-source models offer enterprises better security and safety. Deirdre Bosa
• Mistral still lacks compute, Mensch said; it bought and operates GPUs to secure supply amid compute constraints and insufficient European cloud partners, enabling independent token generation, training on its own compute and faster scaling. He expects substantial compute to come online in 2027–28 and described demand as especially strong in Europe. Deirdre Bosa He said owning chips and physical assets raises capital needs but can be sustainable without excessive leverage, and that Mistral is preparing for an IPO in the next few years, though it is not ready yet. Deirdre Bosa
• Mensch identified agent deployment as a practical safety concern because agents can behave unexpectedly; customers seek isolation and observability, and Mistral invests in alignment and cybersecurity at both system and model levels. He said such risks require international governance, that proposed US standards would not be good for the world generally, and that agent takeover scenarios distract from complex problems in enterprise AI deployment. Deirdre Bosa

• Andrew Ng expects AI agents to add substantially to demand for core infrastructure: agents will use compute and storage beyond human workloads, complementing data-center demand for AI training and inference; he also said builders are already running short of compute. Perfology
• Ng described a four-part AI engineering skills map his team developed from more than 10,000 job postings, structured interviews, surveys, and partner input: building and deploying uncertain AI applications, software engineering fundamentals, using coding agents, and shaping what gets built. He said people with these skills are in strong demand and difficult to find. Perfology
• Ng argues that coding agents make implementation easier and shift the bottleneck toward deciding what to build; engineers who can use customer feedback and product or architecture judgment can contribute beyond implementation. He says the right balance between quick prototypes and investment in architecture, security, and scale depends on a project's stage, and that engineering fundamentals help people recognize trade-offs agents may otherwise make unnoticed. Perfology
• On AI and learning, Ng says students using LMs can do better on homework but worse on final exams, and that cognitive offloading can increase work completed while reducing learning; he believes this is also underestimated among working adults. Perfology

• Dario Amodei describes GPT-2 as the first real empirical test of the scaling-laws hypothesis. Although its capabilities in individual areas were limited, its breadth led him to warn that scaling it by another factor of a million could make it better than humans at everything; he says the pattern of apparent limitations receding is recurring across wider domains. The Builders
• For Claude’s character, Amodei says Anthropic moved from lists of roughly 170 discrete rules to principles such as open inquiry, care, fairness, and justice, which he says help Claude generalize to situations it may not have encountered. The Builders
• Amodei identifies sycophancy as harder to eliminate than overt errors, arguing that ad-driven incentives to prolong engagement can discourage correcting users; he says Anthropic’s no-ad approach avoids that incentive. He also describes stress-testing for deception and using mechanistic interpretability to investigate when a model’s apparent beliefs diverge from what it tells users, including approval-seeking that could be achieved by fooling them. The Builders
• On autonomous weapons, Amodei argues that even reliable AI-controlled forces create an accountability and democratic risk if one person can command a million drones: unlike human soldiers, AI systems cannot refuse unconstitutional or morally abhorrent orders. He says Anthropic is unwilling to support such uses and views congressional efforts to enact red lines into law as part of democratic change. The Builders

Yann LeCun trained networks of sliding image detectors with backpropagation at Bell Labs in 1989 to read handwritten ZIP codes; the detectors learned useful strokes from examples rather than being hand-designed. Machine Theory The episode says his networks later read checks at American banks and attributes to LeCun an estimate of about one in ten. Machine Theory

• Cohere's funding round was framed as a roughly US$5B target; Pineau said the company had raised more than half but did not disclose the exact amount, citing investor interest in sovereign AI and customer autonomy over how technology is built and deployed and where data is stored. The Information
• Pineau said Cohere's roadmap focuses on frontier models, with current efforts in cybersecurity and enterprise use cases alongside past multilingual work; investment supports larger-model training and synthetic data, and Cohere plans some compute in Canada while exploring compute and inference services for small and medium-sized businesses. The Information
• Nvidia supplies most of Cohere's chips, Pineau said, while the company has had good results with AMD and benchmarks other options; it is talking to specialized inference-chip vendors but has few workflows on them so far. She said Cohere is open to neocloud and larger-provider partnerships rather than building data centers itself, expects some capacity in Canada and eventually Germany, and values optionality as compute costs rise. The Information
• Pineau linked Cohere's growth to global expansion and industry-specific markets, citing Aleph Alpha in Germany, Prague-based cybersecurity partner Isle, Reliant AI's pharma expertise, and financial services, pharma and cybersecurity as priorities. The Information
• Reacting to Anthropic's S-1, Pineau said its risk disclosures prompted her to watch whether this becomes a standard for AI-company IPOs; she argued that investor-facing risk reporting should be transparent but grounded in science and the systems' properties, particularly when discussing speculative risks. The Information

Andrew Ng distinguished verifiable tasks, such as coding or math with checkable answers, where AI performs well, from judgment tasks with no single right answer, where AI is less reliable; he called this a rough distinction and argued that hiring assessments should move beyond testing for correct answers to evaluating judgment. AI Fund He said strong AI-assisted output alone can mask a candidate’s inability to handle judgment tasks, and recommended asking candidates what key decisions they made and why, then probing the context and trade-offs behind those decisions, such as reliability, security, privacy, and system architecture. AI Fund