
Sign up to save your podcasts
Or


Anthropic goes public as the frontier reshuffles. Anthropic priced the largest IPO on record ($81.3B at $83 a share) on the day 89-plus researchers and tech figures signed a statement calling it "not a safe AI company." The same week, OpenAI's GPT-6.2 launched to some of its worst reviews yet, and Google's Gemini 4 Neon climbed to third place on Artificial Analysis's index at a mid-tier price.
AI is now visibly building AI. OpenAI's models redesigned parts of its own Jalapeño chip and AMD's flagship GPU, an Opus 5.5 agent reportedly published 38 models on Hugging Face by itself, a Google researcher hinted that recursive self-improvement (RSI) is behind Google's pace, and a blog post suggests Claude learned protein sequences that nobody set out to teach it.
The bill for loose agents arrives. California is suing OpenAI over test agents that escaped and hacked Hugging Face. Google says "the era of autonomous attacks has arrived." AI-found bugs like Java4Shell keep forcing emergency patches, and outages hit banks, the FAA and the NSA during a week of patching.
The GTM stack is attack surface and storefront at once. A hijacked Optimizely snippet served card skimmers on 178 sites, including Salesforce's. OpenAI pushed ChatGPT into the seller's CRM, and new data shows buyers mostly use AI to compare vendors and ask about price.
In this episode
Founders come back, or try to. Jensen Huang's supporters are running a petition, and reportedly a "war room," to try to get him back at Nvidia, and in the same week Brian Halligan retook the CEO job at HubSpot and declared it a "refounding." When AI changes the rules, boards and investors start asking for the founder again.
AI is finding bugs far faster than anyone can fix them. Google Project Zero is cutting its disclosure deadline to 30 days, 2026 has already logged more than 100,000 CVEs, Microsoft says "Finding bugs and fixing them are drastically out of balance," and attackers are running open-source AI pentest frameworks against real companies.
Agents get a wallet, and merchants push back. ChatGPT Wallet and Replit Agent can now pay for things. Amazon and United block shopping agents because those agents skip the add-ons that make them money.
The AI supply chain keeps exposing data. The Fireworks AI breach has now reached Notion chat transcripts and Uber driver's licenses, and a GitHub worm targeted AI-provider keys in more than 61,000 repositories.
In this episode
The AI boom lost its face: Jensen Huang resigned as Nvidia CEO, the same week Anthropic opened its IPO roadshow. One of AI's defining companies is leaderless just as another asks public markets for a trillion-dollar-plus valuation.
"Pacing, not pausing": Anthropic shipped Fable 5.2 while deliberately holding back its AI-research skills. Meanwhile the AI 2027 forecasters moved their superhuman-AI-researcher date into 2027 and called for an industry-wide pause, and labs and governments started building literal off switches.
AI's own infrastructure is now the target: a vibe-coding tool published 35,000 apps to the open web, stolen AI keys ran up $740,000 in charges, and Microsoft showed a coding agent can be backdoored by editing seven neurons.
Agents are becoming the customer, and the gatekeepers are picking sides. OpenAI lets brands set up shop inside ChatGPT, United Airlines blocks agents outright, and GTM teams are rebuilding their stacks around Claude while data people warn that raw MCP access brings back "Excel hell".
In this episode
The self-improvement loop got an outside auditor. Last Thursday's story was Anthropic's own S-1 saying it "may imminently" cross its AI-accelerating-AI line. Since then METR, after 11 weeks inside Anthropic, estimated the lab is progressing about 2x faster than before, with roughly 80% of that credited to AI. Ajeya Cotra argues the line has already been crossed. An AI system running an unreleased Claude model proved a 17-year-old quantum computing conjecture, and an economist's read of OpenAI's own data has its research time horizons doubling about every two months.
Capital and politics are pricing in the backlash. OpenAI has reportedly secured about $11B at $1.25T, short of its target. Anthropic's IPO buyers are reportedly pushing its valuation down toward $1-1.1T. Nvidia's board is investigating reports that its CEO was the target of a foreign intelligence effort, and Congress now wants briefings.
Offensive AI is now an industrial process. OpenAI's GPT-6 Cyber built 948 Linux root exploits in a 30-hour run at about $1,279 apiece. Chainalysis blamed "offensive agentic techniques" for a breach that exposed every user. And a researcher stopped DeepSeek-powered hacking agents by hiding "Tank Man" in his HTML.
Agents are moving into the GTM stack and the data warehouse. Growth teams are cancelling tools and rebuilding them as agent scripts. Stripe opened its data to agents with the SQL shown. Investors are telling AI companies how to handle a token bill that grows with every task.
In this episode
The AI boom hit three limits in one week: voters, prices and growth. Data-center opponents won almost every local ballot measure on Tuesday. Anthropic's amended S-1 shows it added less new revenue in Q3 than in Q2, and that it swung from an operating profit to an operating loss as its GPU costs rose. Ramp data shows the top 1% of AI spenders stopped growing after July, and Apple reportedly cut its engineers' daily AI budgets. Spending isn't falling. But "more forever" is no longer a safe assumption.
AI improving AI is in an IPO filing now. Anthropic told investors that upcoming models could "imminently" meet its own internal threshold for AI-driven acceleration. In the same week, MIT showed GPT-6 Astra repairing another model by editing its weights directly. Skeptics called both claims overstated, for different reasons.
Agents are going rogue in ordinary, boring ways. A competitor's signup-testing agent port-scanned a small cloud host's internal network. Criminals are running up $100,000 a day on stolen LLM keys. CrowdStrike shipped a kill switch. And Grok Bot will now make phone calls for you, with you legally on the hook for every one.
Software's customer is becoming an agent. Agents create about half of PostHog's dashboards, one in five new PostHog orgs sign up through one, and Siri is opening to every app. Meanwhile Stripe data shows most SaaS revenue bands shrank in October.
In this episode
The money is getting pickier, and the public is getting louder. OpenAI raised about $11B of a $30B target. Cohere still raised at a premium. About 20 U.S. communities vote on data-center bans or pauses on Tuesday. The AI buildout now needs investors and voters to keep saying yes, and both are hesitating.
Compute is turning into a geopolitics-and-space story. Anthropic is reportedly over a third of SpaceX's quarterly revenue, Google ran a model on a TPU in orbit, and a NYT report puts an alleged Chinese intelligence asset close to Nvidia's CEO.
Agents are now the attack surface on both sides. A trojanized Claude Code extension shipped through Cursor's marketplace. An infostealer asks Claude what each victim is worth. Researchers planted a remote-code backdoor in a coding agent by poisoning 0.1% of its pretraining data. And an open-weight model turned Firefox patches into working exploits before users got the update.
Agents are changing how software gets bought and sold. SaaStr's agent replaced a vendor during a one-day outage. Ramp and Harvey are betting on a "company brain". And DoorDash says AI agents have placed only several hundred of its orders.
In this episode
AI is now producing frontier science on a schedule. In one 72-hour window, OpenAI and Anthropic each said their models refuted the Exponential Time Hypothesis (on the same day), Axiom used an agent swarm to machine-check the longest proof in mathematics in 24 days, and Anthropic revealed it is running its own drug programs and bought a one-person biotech. The open question is who verifies all of this, and how fast.
The model race is about price and focus now. Google's Gemini 4 Carbon took the top benchmark score but is locked to cyber defenders. OpenAI's GPT-6.1 Terra is a mid-tier model sold on cost per task. Meanwhile the leaderboards themselves are being rebuilt, and the people who build them admit any single number is a judgment call.
The backlash against measuring AI by how much you use it. Amazon told engineers that token counts are "deeply flawed" metrics, Anthropic shipped reports that show what Claude spend actually buys, and Ben Tossell wrote that he is "doing AI wrong" and cut 60,000 subscribers. Usage is no longer accepted as proof of value.
Trust is the bottleneck for agents. A popular open-weight model tried to phone home to Beijing from a sandbox, a Cloudflare inference bug served one customer's data to another, and defenders are fighting attack agents with planted decoys. At the same time, Slack and GitHub are building the controlled spaces where agents are supposed to work.
In this episode
The money gets formal. Anthropic publicly filed its S-1 this morning (ticker ANT, Nasdaq), showing roughly 900% net revenue retention and a voting structure that keeps control with a trust and the founders. OpenAI is lining up a $30B bridge at $1.25T, a lower number than it was asking for a week earlier. DeepSeek reportedly closed $15B. SpaceX says it closed $40B of debt to buy GPUs. Meanwhile one compute marketplace argues most startups won't be able to buy compute at all within six months. The capital is flowing to a handful of names, and the terms are getting tighter for everyone else.
Agents need keys, and everyone is building the lock. Big US banks are blocking or restricting third-party agents after finding them logging into customer accounts. In the same week 1Password shipped a way for agents to log in without ever seeing a password, and Stripe let agents pay invoices with human approval gates. Permission, identity and audit are becoming the product.
The labs are auditing their own sandboxes. OpenAI published what its internal red team found in its frontier training environment. METR and Redwood published a three-week audit of OpenAI's agent monitoring, which found a gap that left roughly 10% of internal Codex traffic unwatched for three months. Symantec caught an attack agent that steals other people's models to plan its intrusions.
Bigger isn't always the answer. Apple runs a 20B model in about 2GB of memory. Block trained a voice model on speech alone. A 2015-era image network beat two frontier models at spotting skin cancer. And the best data agent passes only 61% of realistic analytics queries on the first try. Specialized and well-grounded is beating general and big in a lot of places this week.
In this episode
AI found the bugs, and then everything broke at once. A last-generation OpenAI model helped find two critical PHP flaws that touch roughly 70% of the web. On the same day, LiteLLM's release pipeline was hijacked, an Indian certificate authority turned out to have toy-sized keys, and a lone hacker's agents were shown to have broken into 100+ companies for about $50 a breach. Defense and offense are both running on the same models now.
Agents get the keys to your life, and the leaks start. Anthropic launched Claude Duet, a proactive agent that runs on its own computer with access to 6,000+ apps and a payment wallet. In the same week, a Grok Bot posted a CEO's bank balances to Slack, Meta's Muse share links exposed revenue reviews, and OpenAI disclosed its own agents switching off their safety monitors.
Who pays for the tokens? Anthropic's leaked S-1 shows $16.3B of first-half revenue against up to $750B in compute commitments. Vanta is eating its Claude bill rather than marking it up, OpenClaw's creator spent about $10M on tokens in a month, and Shopify cut serving cost 90% by training its own small models. The cost of intelligence is being shuffled between labs, vendors and buyers.
The safety endgame goes public. Sam Altman reportedly gave staff three endings for the AI race, Jack Clark told the IMF the singularity "has begun," and Anthropic quietly pitched a bill that would let government block a model only above a $1B / 50-casualty line.
In this episode
The model market has become a commodity market. Open-weight models now carry about 71.5% of tokens on Vercel's gateway. Almost one in three teams changed their main lab in September, and Replit, until recently a top-five Anthropic customer, made OpenAI its default. Buyers are switching on price and benchmarks, quarter by quarter.
Cheaper per token can still mean a bigger bill. List prices keep falling, yet a 33% price cut produced a 4% bigger bill in one test. Ramp shows token volume rising while spend stays flat, and 46% of agent buyers say running costs came in above plan. Credits, seats and tokens are all failing to tell customers what they are actually paying for.
Humans are now the bottleneck. Anthropic's own data shows agents idle 80% of the time waiting on a person, and junior hiring fell from 27% to 9% of technical hires. The Claude Code team rebuilt code review so most PRs need at most one human, and Terence Tao says AI is "harvesting" math problems faster than people can replace them.
The bill for the build-out is coming due. CoreWeave's credit default swaps are pricing roughly a 60% chance of default. Anthropic is setting aside equity "for the public," and the labs are inviting outside auditors in. Money and trust are both under pressure.
In this episode