• In the AMI Labs exchange, LeCun described a French company headquartered in Paris, with major offices in Paris and New York and smaller offices in Montreal and Singapore; it had begun operations in March, employed about 65 people, and raised about $1 billion in seed funding, which he believed was the largest seed round outside the U.S. Evident Insights He said the company anticipated some industry successes over the following year or so. Evident Insights
• The company’s thesis is physical-world AI based on learned world models: predict the world state after an action, then plan action sequences, using hierarchical subgoals for longer and uncertain tasks. Evident Insights LeCun’s long-term proof point is a humanoid robot that can handle tasks a 10-year-old can do without extensive explicit training or thousands of demonstrations; he argues LLMs are not capable of this kind of planning. Evident Insights
• LeCun described a non-generative joint-embedding predictive approach that predicts in abstract representation space and ignores unpredictable details; he said this is relevant to financial data, which he characterized as mostly noise and unsuitable for prediction of a single specific outcome. Evident Insights
• LeCun said much of the visible research on world models and joint-embedding approaches is coming from Asia, particularly China, which he sees catching up faster than Silicon Valley; he attributed the regional focus to strength in manufacturing and robotics and said Silicon Valley remains heavily focused on LLMs. Evident Insights

• Alpha says its learning engine combines cognitive and learning science with LLMs; it personalizes lessons to keep students around 80–85% correct, adding scaffolding when work is too hard and progressing students faster when it is too easy. The school claims students learn twice as much in two academic hours as in six classroom hours plus homework. a16z
• Alpha reports growing from Austin to 50 campuses after adding 12 the previous year, and describes demand as high. a16z
• Alpha’s Founder School, opened in August, is for high-school students who want to build businesses. It guarantees a student will start a business earning $1 million in profit during high school or receive tuition back, with milestones from $100 through $1 million; the founder said it was only a couple of months into the process and students were already moving quickly. a16z
• Alpha says more than 70 students from MIT, Stanford, and UT interned there over the summer and worked on a proposed educational LLM intended to be free globally; the stated rationale was that each sovereign would need its own educational model. a16z

• LeCun left Meta to pursue world-model research through a new company after identifying industrial process control, manufacturing and robotics as its main applications; he said Meta had de-emphasized robotics and refocused on LLMs around spring 2025, diverging from his research agenda. TBS CROSS DIG with Bloomberg He says he raised enough funding for a serious effort and is spending heavily on GPUs to train video and other world models. TBS CROSS DIG with Bloomberg The interviewer identifies Toyota Ventures as one of the venture’s early investors. TBS CROSS DIG with Bloomberg
• JEPA predicts in abstract representations rather than generating pixels; LeCun says it has worked for self-supervised vision and video, and describes a world model as predicting the consequences of actions to support planning. TBS CROSS DIG with Bloomberg He sees LLMs as useful for distilling text knowledge but insufficient for real-world understanding and planning without world models. TBS CROSS DIG with Bloomberg He says many teams that tried VLM/VLA approaches for robotics are shifting to world models, which he describes as a hot research topic not yet widely adopted in industry. TBS CROSS DIG with Bloomberg
• Initial applications are complex industrial systems: sensor, video, camera, vibration or audio data could support future-state prediction, anomaly detection, predictive maintenance and optimal control; LeCun sees manufacturing as an initial market, with autonomous driving and domestic robots longer-term. TBS CROSS DIG with Bloomberg As a readiness caveat, he says humanoid builders do not yet know how to make robots useful and are betting on AI progress over the next two to three years. TBS CROSS DIG with Bloomberg
• LeCun advocates free, open frontier foundation models as a base for vertical applications and localized assistants; he warns that concentration among a few AI companies could make people’s information diets—and culture—more uniform. TBS CROSS DIG with Bloomberg

• Helix Carbon is developing a containerized electrochemical retrofit for blast furnaces: it says the system can process off-gas without carbon capture or gas separation, convert CO₂ to CO, and return it to the furnace to displace some met coke while retaining existing blast-furnace assets. The company describes those assets as multibillion-dollar installations that can operate for over a century. TechCrunch
• At the presentation, its demonstrated system was a 0.5-kW bench unit fed by cylinders mimicking blast-furnace gas. TechCrunch Helix planned a 3-kW pilot with a Minnesota research furnace in Q1 of the following year, followed about 12 months later by a 150-kW commercial unit; it said the commercial unit was already under purchase order with a U.S. Fortune 500 company. TechCrunch
• Helix estimates savings of $14 per ton without green premiums or incentives. Its modeled electricity price is $0.05/kWh, with economics becoming less favorable around $0.07/kWh; the company says it can pause operations during price spikes to manage gas costs. TechCrunch
• The planned model is hardware sales as customer capex plus annual maintenance, with small deployments expanded over time at each furnace. TechCrunch CEO Evan Hos cites work on iron-and-steel decarbonization policy with the Biden administration; the company described a team with three technical PhDs and engineering expertise, while acknowledging it lacked iron-and-steel insiders and was adding industry advisers in connection with its seed round. TechCrunch Helix says the carbon loop is patented and some electrolyzer internals are trade secrets. TechCrunch

• Sam Altman said GPT-6, launched in August, had OpenAI’s strongest adoption of any new model; he described an unusually rapid recent pace of progress and cited gains in difficult math, coding, enterprise software, security, and science. CNBC Television
• Oracle’s event host said more than 70% of Oracle’s software code was being written by AI, with the share increasing. Altman said company-wide AI rollouts were prompting employees who had not considered themselves software creators to build apps; speakers also cited workers’ self-reports of feeling twice as productive, rather than a measured productivity result. CNBC Television
• Oracle launched AI Security Mesh to analyze systems, validate attack paths, and help produce and deploy fixes. Altman urged companies to prioritize understanding and mitigating security threats, saying AI is exposing existing vulnerabilities rather than creating the underlying bugs; he argued that addressing them at scale will require automation and new remediation approaches. CNBC Television
• Altman said AI is helping discover science and invent technology; he also said AI helped design Jalapeno, an inference chip being deployed in OCI that he described as the world’s best, and accelerated its development. CNBC Television
• Oracle said it was bundling startup-focused financial software with OCI and OpenAI credits. CNBC Television

• Moritz pitches CrateDB as one engine for transactions, analytics and AI on live data, intended to avoid separate databases and synchronization delays. TechCrunch He describes the team as five database PhDs and says they rebuilt the database components from scratch while reimplementing PostgreSQL’s API; he cites PostgreSQL compatibility as a way to lower replacement friction. TechCrunch He announced a fully managed cloud service and demoed an I60 instance at about $2,500/month. TechCrunch The pitch identifies the company as CrateDB but also calls the product and cloud launch “SingleStore DB” and “SingleStore DB Cloud,” leaving a material naming inconsistency unresolved. TechCrunch
• The founder contrasts CrateDB’s single-engine approach with Databricks’ two engines joined through common storage, claiming the latter can still have a 2–5 minute delay; he treats Databricks’ launch as validation that customers want this category. TechCrunch The pitch claims a lower-tier instance can process over 60,000 transactions per second while complex analytical queries return in under 100 ms. TechCrunch The founder also cites one customer/POC reporting 10× performance at one-tenth the cost versus an analytics-focused database cluster, using one server. TechCrunch
• The founder reports over 2,000 on-premises projects started in the prior month, downloads growing over 20% monthly to more than 2,000 in September, and hundreds of users asking for a hosted service. TechCrunch He says adoption to date was mostly on the Community Edition and describes use cases generally, clarifying that the stock-exchange example was not a literal exchange. TechCrunch The target buyers are organizations for which real-time analytics is a hard requirement, such as for competitive or regulatory reasons. TechCrunch

• LeCun argues that current LLM-based and generative systems handle symbols better than the high-dimensional, noisy data from sensors and video, and do not anticipate the consequences of physical actions; he says text-only training will not produce human-level or physical intelligence, which depends on adapting to new situations with little prior training ビジネス＋ITチャンネル.
• His proposed world-model approach predicts the state resulting from candidate actions, then plans by optimizing for task completion subject to safety objectives; he describes this as a blueprint for intrinsically safe agents, rather than the unconstrained outputs of LLM-based agents ビジネス＋ITチャンネル. JEPA avoids pixel-level reconstruction and instead predicts in an abstract representation space; LeCun says such learned models could apply to complex systems including jet engines, chemical plants, power plants, patients, and cells ビジネス＋ITチャンネル.
• LeCun identifies Japan’s precision manufacturing and high-quality industrial data as advantages for physical AI, and urges the country to move quickly as a first mover ビジネス＋ITチャンネル.
• Project Tapestry aims to build an open global AI foundation that countries and regions can contribute to; participants keep their training data and share model parameter vectors rather than sending their data to the consortium ビジネス＋ITチャンネル.

• Altman said GPT-6 was a step-change in model capability and OpenAI’s strongest new-model adoption to date, with use spanning coding, enterprise software, security and science. CNBC Television
• Altman urged companies to prioritize understanding and mitigating AI-related security threats: he said AI is exposing longstanding vulnerabilities, while patching every bug across the supply chain may be impractical, creating a need for services that continuously monitor traffic and prevent exploitation. CNBC Television
• Altman said AI is already helping with scientific discovery and chip design; he cited a chip called Jalapeno being deployed in OCI, described it as the best AI inference chip, and said AI accelerated its development. CNBC Television
• Broad enterprise AI rollouts are prompting people who do not consider themselves programmers to build apps and automate tasks, lowering the barrier to software creation. CNBC Television

• Niobia AI is building a factory troubleshooting system: Lab OS fine-tunes open-source models on manufacturing processes and links plant information, while its Edge product connects directly to industrial equipment to diagnose and troubleshoot defects TechCrunch. The company reports resolving an injection-molding defect that previously took two weeks in three hours; its current “level 2” system detects defects and recommends fixes, with human approval required before applying them TechCrunch. It says competitors often stop at detection, whereas Niobia integrates with equipment to fix defects on the line TechCrunch.
• Niobia says it prices at 2% of production-line CAPEX and is starting with thin-film coatings, which it estimates as an $8B market; it cites $200B for advanced manufacturing and $800B for all manufacturing as expansion opportunities TechCrunch. It reports 32 companies in its pipeline, three active pilots, four signed LOIs, $65K in locked revenue, and over $600K in qualified pipeline, which it aims to reach by next year TechCrunch. One early customer is a thin-film-coating equipment manufacturer seeking to connect inline process parameters to final-product outcomes and improve recipes TechCrunch.
• For customers it has interviewed, equipment may not be internet-connected and they do not want data leaving the facility; Niobia offers Edge deployment and says it connects to industrial PLCs and other manufacturing data sources TechCrunch. It says its small language models are fine-tuned for specific processes, and targets full autonomy for coating tools by mid-next year TechCrunch. The company says it started about a year before the pitch and has a five-person team, including four PhDs, with 50 years of combined deep-tech experience TechCrunch.

• Industrial Next offers AI for dexterous factory manipulation, bundled with compute and sensing for existing robots or a compatible work cell; its stated recurring revenue comes from software licensing, with deployments expanded task-by-task where the economics make sense. TechCrunch
• Co-founder Lucas Pankau said the system is deployed in consumer-electronics factories and has produced more than 7 million units of a popular smartphone. He also described a station producing a previously unseen, unannounced phone model 60 seconds after it was turned on; he said those phone operations run at roughly a three-second cycle time. TechCrunch
• Industrial Next announced a first partnership with BMW Group, beginning with a few tasks in an arena to evaluate its dexterous-manipulation technology for manufacturing. TechCrunch
• The company says it developed its own foundation model because available models’ training data was not manufacturing-specific enough for repeatable, high-precision performance at factory speed. Its demo was about 90% simulation-trained and post-trained with roughly two hours of real-world data. TechCrunch
• The team’s cited experience includes Pankau’s Tesla Autopilot and manufacturing work and Waymo sensing and compute development, co-founder Allen’s GM self-driving-car work, and CTO Gil Louie’s Meta robot-manipulation work and stints at MIT, CMU, and the University of Washington. TechCrunch
• The company is initially focused on consumer electronics and automotive manufacturing; it identified medical devices as an adjacent opportunity but said it has not yet solved AI-system certification for that sector. TechCrunch

• Perplexity is positioning its agents as an operating system for work: its personal-finance use cases include connecting a portfolio, tracking spending and budgets, and helping users construct portfolios or investment strategies, while users retain control of trades and purchases. Srinivas said recurring automation accounts for about one-third of usage in these use cases. New York Stock Exchange
• Srinivas argued that users want AI to research and help decide what to buy more than to execute purchases; he positioned accurate search, personalized comparisons, citations, and auditability as the valuable—and trust-building—part of the experience. New York Stock Exchange
• Perplexity's stated trust approach includes tracing financial figures to sources, using secure connectors such as Plaid without storing users' personal information in logs, and running agents in secure sandboxes. Srinivas said the company had open-sourced its Space sandbox after testing rogue-agent behavior and adding guardrails, and that neo-cloud providers were adopting it. New York Stock Exchange
• Perplexity says it aims to be a model-neutral orchestrator, choosing the best model for each task rather than steering users toward a model it owns; Srinivas described model providers as supply-chain contributors who receive feedback on how users respond to their releases. New York Stock Exchange
• Srinivas identified calibrated uncertainty and continual learning—gathering missing information, asking useful follow-up questions, and revising answers—as capabilities that do not yet exist but could let AI address unknowns. New York Stock Exchange

• Chronosphere, founded by former Uber engineering leaders Martin Mau and Rob Skillington, received Greylock’s first check in 2019 and was acquired by Palo Alto Networks for $3.35 billion in 2026, according to the video. Greylock
• As observability leaders consolidated metrics, traces, and logs into broader platforms, Chronosphere began building a logging product in 2021, aiming to get ahead of the market by two to three years. Greylock By 2024, ARR growth that had once topped 150% had fallen below 40% for two years, and logging had yet to produce a major commercial win. Greylock
• After missing its year-end target, the company weighed a down or structured round against betting on two large deals; the founders chose to keep executing, with Greylock helping secure investors’ support for more time. Greylock Two quarters later, logging landed its first major customer and Chronosphere signed foundation-model clients; a further frontier AI lab migrated from its existing vendor at roughly half to a third of the cost with better performance. Greylock AI companies generated more data and faced higher cost pressure, while existing tools were less effective; Chronosphere’s product, initially built for a different environment, fit those needs, and the company reported its fastest-ever growth in 2025 after expanding from cloud-native to AI-native customers. Greylock

• Prism, led by founder and CEO Martha Sila Underwood, sells bank-embedded inheritance infrastructure: Designate lets customers update beneficiary data in their bank app, routes submissions to bank staff, and surfaces stale or missing designations; Anchor handles documents and validation, while Continuity is intended for money movement. TechCrunch Its stated wedge is connecting banks with beneficiaries before wealth moves, with the aim of retaining the relationship and deposits; the workflow captures beneficiary contact information and account-holder consent for early engagement. TechCrunch
• Prism reports a live Mississippi bank with $37B in assets, integration into all three Jack Henry cores, three patents in process, 39 active pipeline institutions, 1,700 institutions reachable through Jack Henry, and 12-week deployments. TechCrunch Underwood later said the company was working with two banks and had several credit-union contracts out; its stated target is banks and credit unions under $30B, so the disclosed $37B live bank is above that target. TechCrunch The company charges an annual fee based on institution size plus a transaction fee for accounts it helps keep in-house. TechCrunch
• Underwood frames the market around $124T expected to shift to the next generation and cites a September bank-director survey saying 70% of financial institutions were not ready to capture next-generation account holders. TechCrunch She reports 25 years at companies including IBM, BBVA, PNC, and SAP; she describes her team as having decades of bank operations and enterprise-technology experience. TechCrunch
• A key validation caveat: no customer had yet died on the platform, so the inheritance money movement had not been exercised. TechCrunch Underwood identifies bank-core providers as well-positioned potential entrants, but says they are not pursuing this now due to other priorities and the technical lift, and are instead integrated with Prism. TechCrunch

• LNK Energy is commercializing an MIT-developed LOHC hydrogen-delivery system: a room-temperature liquid designed to move hydrogen using existing diesel tankers, pipelines, and gas stations, with a catalytic release unit supplying a fuel cell or engine. TechCrunch The company claims delivery costs fall from 95% of fuel-station price to 6%, and says the resulting hydrogen can cost 26% less than diesel. TechCrunch
• LNK is targeting Los Angeles-area drayage trucking and says it is working with fuel-cell maker Symbio toward a commercial truck operating in California within 18 months. TechCrunch Its model is to sell hardware kits to partners and earn margin on fuel supply; the company is raising a seed round. TechCrunch
• The company reports 50 hours of testing and a mini-fridge-sized release unit, compared with prior systems it describes as shipping-container-sized; it says the technology is patented through MIT. TechCrunch The reactor originated in the CEO’s MIT PhD work, and the CTO joined the project as a researcher to develop the powertrain; the company identifies MIT Energy Initiative director William H. Green as CSO. TechCrunch
• Commercialization remains contingent on execution: LNK says it does not yet have permits, although agency discussions have raised no red flags. TechCrunch It lists UC Riverside third-party validation and CARB licensing among upcoming milestones, and identifies engineering, manufacturing, and reaching viable prices at low initial volumes as remaining challenges. TechCrunch

• Flocker pitches a customer-data platform that maintains itself for AI agents, routing across 80+ data vendors and using agents to build and maintain data-cleaning pipelines and a customer ontology that connects business records. TechCrunch
• The company reports hundreds of customers, including Perplexity, AngelList and Wise, and says customers have generated millions of dollars in pipeline and saved over 70% of customer-research time; it also reported 14% compound month-over-month growth in recent months. TechCrunch Annual contracts start around $10,000; the founder said some accounts, including Wise, Perplexity and AngelList, expanded 20x and now pay six figures. Flocker sells initially to go-to-market teams and expands into data engineering. TechCrunch
• The founders traced the company’s origins to a consumer Chrome extension in 2024; the team is eight people in Toronto and includes former customers. TechCrunch Flocker says it is going open source, while the founder argues its moat is the agent’s context for choosing which data provider to use for each company and use case. Judges questioned the crowded market and whether Anthropic or ChatGPT could replicate the capability; the founder named Salesforce Data 360 and in-house Snowflake databases as alternatives. TechCrunch For complex fintech data, Flocker says customers can keep data inside their own ecosystem and that agent training may be needed; general SaaS onboarding is mostly automated. TechCrunch

• Andrew Ng is bullish on databases and core infrastructure: he expects agent use to add compute and storage demand on top of AI training and inference, and says builders are already running short of compute. Perfology
• DeepLearning.AI’s AI-engineering skills map drew on more than 10,000 job postings plus interviews and surveys; it identifies building and deploying AI applications, software-engineering fundamentals, coding-agent use, and shaping what gets built as core skills. Ng says people with these skills are in great demand and hard to find. Perfology
• Ng says coding agents make building easier, shifting the bottleneck toward deciding what to build; engineers who can bring product judgment and customer feedback into development can contribute more. For early-stage projects, he favors safe, quick prototypes and rapid learning, while architecture and security deserve more attention as project needs and risks change. Perfology

• Longshot Space CEO/cofounder Mike Grace said the company had raised about $20 million and was raising a Series A, with customer LOIs and about $15 million in Department of War hypersonic-technology-development bookings. His demand thesis is that Elon Musk will use launch capacity for his own payloads rather than other companies’ satellites. TechCrunch Grace cited his NASA experience and said his cofounder had built rockets and satellites and worked at “Neurolink.” TechCrunch
• Longshot proposes a ground-based multi-injection gun to accelerate a hypersonic lifting body, followed by a solid-rocket second stage to reach orbit. TechCrunch Grace showed a demonstrator at the Barstow test facility and said a 30-inch-interior-diameter gun would be operational in December; he also described a roughly half-kilometer system as a summer goal. TechCrunch
• The stated MVP is intended to compete with Starship on unit economics, with a 15-ton vehicle carrying a 2-ton payload; Grace targets reducing acceleration to about 200 g by lengthening the accelerator. TechCrunch He sees dense semiconductors, solar panels, GPUs, and telecom or space-compute equipment as suitable payloads, but said packaging and orientation would need adaptation; he considers large optical instruments and 20-ton geostationary satellites poor fits for the system. TechCrunch
• Grace said the accelerator works, but identified heat-shield survival and navigation through the lower atmosphere at Mach 14–16 as challenges still to solve. TechCrunch

• Co-founder Brad Joffe framed Apate AI’s opportunity around global scam losses he said exceed $1 trillion annually. TechCrunch Apate deploys persona-based conversational AI across voice, text, email and web to engage scammers, waste their time and extract intelligence such as bank accounts, crypto wallets, malicious URLs, scam scripts and attack infrastructure for defensive use. TechCrunch
• Apate reports more than 9.3 million autonomous conversations, the equivalent of 17 years of scammers’ time wasted and over $33 million in avoided scam losses; cited customer outcomes include a 76% reduction in scam losses at a major Australian bank, $183,000 saved by disrupting a live campaign in one afternoon, and mule accounts identified up to 61 days earlier than existing detection. TechCrunch
• The company says it is in full production with major banks, telcos and government agencies in Australia, Singapore, South Africa, Ireland, the US and UK; Commonwealth Bank of Australia and Vodafone were design partners. TechCrunch Customers can start with a low-integration online portal or use APIs, and providing scam leads can make investigations more powerful; a product that has an AI agent contact customers when a bank flags a potentially risky transfer is in pilot. TechCrunch
• Joffe said adoption is harder in geographies without a strong immediate incentive to act; he pointed to UK reimbursement requirements and said scam prevention is receiving greater attention in the UK, Australia and Singapore. TechCrunch
• The pitch identifies co-founder Professor Dali Kaafar as a cybersecurity researcher named Australian AI Researcher of the Year, Pete Ackerman as an experienced government, technology and finance executive, and Joffe as a lawyer who says he spent 10 years scaling and exiting technology companies. TechCrunch

Plug and Play’s second Albanian batch graduated 21 startups; the update said 10 would go to Silicon Valley and more than four investments would be made in the batch. Plug and Play Tech Center The program reported that all cohort companies had initial traction through pilots or first implementations, with some already signing large contracts. Plug and Play Tech Center One founder said the accelerator helped clarify product-market fit after sales meetings and pitches exposed that they had not yet figured out the problem; they said the fit was now clear. Plug and Play Tech Center

• Product and team: Satellite founder and CEO Ramu Ocheng Ofulo says he is formerly of SpaceX, Google Cloud, and Tesla. TechCrunch The company is building a software layer to deploy AI and applications across compatible spacecraft: Sat AI analyzes data onboard, SatNet moves results over intermittent links, and Sat Cloud orchestrates workloads; operators retain command authority and Satellite says it does not handle flight controls. TechCrunch CTO Nelson Synjen is described as one of the first people working with NASA on distributed computing in space, and Sharon as a former ISS spacecraft engineer. TechCrunch
• Traction and financing: Satellite reports more than $500,000 in commercial contracts and an $8 million seed round. TechCrunch Its reported on-orbit work includes an uplink to a Rogue Space satellite that reduced telemetry and log data by 97%, and preloaded software on a Momentus satellite running Google Gemma diagnostics and cybersecurity health monitoring. TechCrunch The company says it is “in line to win” a $1.3 million NASA SBIR Phase II after Phase I, so the Phase II funding is prospective, not a secured award. TechCrunch
• Business model and scaling: Satellite charges operators $4,200 per spacecraft per month and software vendors $10,000 plus usage fees; it pitches operational efficiency for operators and the ability to sell satellite resources to other users. TechCrunch The stated plan is to start with efficiency on individual spacecraft, then cross-link heterogeneous satellite systems; its target is 15 nodes this year. TechCrunch It is working with NASA on in-orbit standards and CCSDS on store-and-forward standardization, while positioning itself against closed, vertically integrated orbital-computing efforts including SpaceX and Google. TechCrunch The Q&A acknowledged that standardization could also benefit competitors. TechCrunch