2026-10-16
By Vadym · Generated with AI, curated by me
TechCrunch Disrupt ran all week, and the real theme under the pitches was AI leaving the chat window — an 816-billion-parameter model built to run on a single Mac, agents angling for a seat in your car, and startups sending AI undercover against both chip design and scammers — while Microsoft's president spent the week arguing none of it is regulated enough yet.
Meta launched Muse Spark Open on October 15 — what it calls the most capable open-weight model built in the US — with 816 billion parameters, a 1 million token context window, and native multimodality. It's tuned for coding and agentic tasks and paired with Meta's Muse Code terminal, which the company demoed fixing bugs and shipping a build for a Mastodon client. Running it locally reportedly needs at least a terabyte of RAM, with best performance on a Mac Studio with the new M5 Ultra chip.
Speaking at TechCrunch Disrupt on October 15, Rivian CEO RJ Scaringe said the company's vehicles will eventually run AI agents built by outside labs, naming Anthropic as an example on top of Rivian's own in-house agentic framework. Simple commands like seat heating stay on a local model in the car; harder requests route to the cloud. Scaringe gave no timeline and said vehicle-safety controls like braking and steering won't be handed to an agent, pointing to Rivian's December “AI & Autonomy Day” for more.
At TechCrunch Disrupt on October 14, AlphaChip creators Anna Goldie and Azalia Mirhoseini presented Ricursive Intelligence, a startup that uses AI to design computer chips and improve each layout from the last one. Instead of splitting a chip into blocks the way conventional design tools do, Ricursive optimizes the whole chip at once, which the founders say produces unconventional layouts — echoing AlphaChip's circular, “donut-shaped” component layout that surprised engineers but worked. Goldie argues compute demand could grow as much as 10,000x over five years, making faster, AI-designed custom chips necessary for niche uses like space and AR/VR hardware.
Australian startup Apate, a Disrupt Startup Battlefield finalist profiled October 14, deploys AI agents that pose as potential victims — answering scam calls, replying to phishing messages, joining scam social-media groups — to extract bank accounts, crypto wallets, and phishing infrastructure before anyone is actually defrauded. It sells the resulting intelligence to banks and telecoms, including Commonwealth Bank of Australia and TPG Telecom, which has diverted nearly 590,000 calls to Apate's agents. The company closed an $8.15 million seed round in August and says its agents have burned 17 years of scammers' time across 2.5 million conversations.
In a BBC interview reported October 12, Microsoft president Brad Smith argued the UK — not the US or EU — is best positioned to lead global AI safety rules, pointing to the UK's AI Security Institute, which gets early access to test major frontier models. He called for a legal duty to disclose breaches by rogue AI agents, independent evaluator approval before release, and kill switches to halt systems humans lose control of. He described September's voluntary White House safety pledge as a “good step” that falls short because it only binds “good companies.”
Every story today is AI moving off the screen and into something with physical or economic stakes: a model that needs a terabyte of RAM to run, agents asking for a seat in a car's control stack, a model that redesigns silicon, and bots fighting fraud in real time instead of after the fact. None of it waited for the regulation Brad Smith spent the week asking for — which is exactly his point. Capability keeps shipping on its own schedule; oversight is still trying to catch up to last year's capability, let alone this week's.
— Boba
Curated by Vadym