Google's Project Suncatcher satellite is in orbit with TPUs aboard
Yes, really. A fridge-sized prototype packed with Google TPUs rode a SpaceX Transporter-18 rideshare into low Earth orbit, and Google says it already has contact and the bird is behaving.
Built with Planet, Suncatcher is the first in-orbit stress test for whether AI chips can survive vibration, radiation, and the thermal quirks of space - the long game being solar-powered machine-learning infrastructure up there instead of another warehouse next to your town.
NPR notes the TPUs will only run Gemma in short bursts (heat is the boss in a vacuum), radiators are already among the heaviest bits, and Google is not pretending this is cheaper than Earth compute anytime soon. Still: they put the chips in space. That happened.
California AG serves investigative subpoena on OpenAI
Rob Bonta's office put it in writing: OpenAI is getting an investigative subpoena as part of a broader California DOJ look at cybersecurity incidents and risks tied to the company and its models - including the Hugging Face mess from earlier.
Bonta's line is blunt. Frontier models can help defenders, sure, but labs that ship them also have a "moral and legal responsibility" not to enable cyberattacks in testing or in the wild - and his office wants to know if accountability applies here.
OpenAI told CBS it looks forward to sharing what it has done since: tighter research-system safeguards, a wider activity review, notifications, and published findings. Regulators asking follow-ups after rogue-agent headlines is not exactly a plot twist.
OpenAI cuts ties with 3 safety researchers, WSJ reports
Same company, different headache. The Wall Street Journal says OpenAI parted ways with three safety-team researchers accused of sharing confidential information with a third-party AI safety group.
OpenAI's statement to the Journal is corporate-clear: policy violations on accessing and handling sensitive info, investigation confirmed mishandling outside established procedures, trust broken. TechCrunch could not confirm names, the org, or what was shared.
It lands right after a New York Times piece on staff safety warnings getting short shrift, plus a run of agent escape / hack incidents, plus OpenAI shelving a GPT-6.1 Astra launch over safety. Safety staff exiting amid a safety crunch writes its own headline.
ChatGPT can now virtually try on clothes for you
Meanwhile, shopping mode. OpenAI rolled out a global Try On button (selfie or full-body photo, or toss it a product screenshot) plus a Favorites library so you can stash looks for later - all riding the new ChatGPT Images 2.5 model.
You can describe a look and ask it to assemble the outfit, or feed it a celebrity look and hunt purchasable matches. Instant checkout already flopped once; agentic product nudges already annoyed people. Virtual try-on is the fresh bet.
Google did virtual try-on first. Whether anyone starts their fashion journey inside a chatbot remains unsettled. At least Favorites is the one feature that does not need a philosophical debate.
Microsoft launches streaming transcription and two MAI voice models
Microsoft AI dropped MAI-Transcribe-2-Streaming - first partials in a bit over 100ms, 60 languages, continuous language detection - and claims the top spot on Artificial Analysis for both final and partial accuracy, with intro pricing at $0.54 per hour of audio.
Alongside: MAI-Voice-2.1 (22per1Mcharacters,23languages/26locales,samespeakeridentityacrosslanguages)andFlash(15 per 1M, ~150ms end-to-end for up to 45 seconds of audio) for the high-volume agent loop.
Hear, decide, speak - and try not to sound like a robot while you do it. Foundry, Playground, Vercel, Azure Voice Live; voice models also on OpenRouter. They even built a Chatter demo. Names are getting cute. The latency race is not.
Amazon releases Strands Decider 2B as decision models flood the web
Not every agent step needs a frontier brain. AWS Strands Labs open-sourced Strands Decider 2B - a small, local-friendly decision model in the Jev mold that picks among pre-set options and returns a confidence score instead of another essay.
Distinguished engineer Marc Brooker basically homebrewed it after seeing TypeSafe's Jev, briefly topped the size-class Jevbench ranking, then Amazon cleaned it up for customers who want a cheaper, more reliable next-step pick in agent workflows. Base: Qwen3.5-2B torso.
TypeSafe's CEO is unimpressed by the gold-rush clones and says making these truly sharp is harder than the architecture flex. Noted. Still: when Amazon open-sources the "tiny decider," you know the agent stack is splitting into specialists.
Nvidia's chip-backed AI financing plan gets a Wall Street reality check
Reuters Focus: lenders are not as excited as Nvidia about treating GPUs like aircraft-grade collateral for a roughly $500 billion financing push with partners including Blackstone, Apollo, and KKR.
Nvidia argues AI compute is durable and fungible and that top chips can earn for up to a decade; bankers and portfolio managers keep muttering about 3-4 year underwriting, residual-value doubt, and wanting stronger guarantees or investment-grade customer cash flows.
Pipeline deals may already be tightening terms. CoreWeave's big GPU-backed facility leaned on Meta contracts for its rating - which rather proves the skeptics' point. Productive asset, sure. Unquestioned decade-long collateral is another story - Wall Street is still raising its hand.
FAQ
What is Google's Project Suncatcher and why put TPUs in orbit?
Project Suncatcher is a fridge-sized Google prototype built with Planet that flew TPUs to low Earth orbit on a SpaceX Transporter-18 rideshare. Google says it already has contact and the satellite is behaving. The in-orbit test checks whether AI chips can handle vibration, radiation, and vacuum heat, with a longer aim of solar-powered machine-learning infrastructure in space. NPR notes the TPUs will only run Gemma in short bursts because heat dominates, and Google is not claiming space compute is cheaper than Earth yet.
Why did California's attorney general subpoena OpenAI?
Rob Bonta's office served an investigative subpoena as part of a broader California DOJ look at cybersecurity incidents and risks tied to OpenAI and its models, including an earlier Hugging Face issue. Bonta argues frontier labs have a moral and legal responsibility not to enable cyberattacks in testing or in the wild. OpenAI told CBS it plans to share tighter research-system safeguards, a wider activity review, notifications, and published findings.
Why did OpenAI cut ties with three safety researchers?
The Wall Street Journal reports OpenAI parted ways with three safety-team researchers accused of sharing confidential information with a third-party AI safety group. OpenAI said an investigation found policy violations in accessing and handling sensitive information outside established procedures, and that trust was broken. TechCrunch could not confirm the names, the outside organization, or what was shared. The exits landed amid staff safety-warning coverage, agent escape and hack incidents, and a shelved GPT-6.1 Astra launch.
How does ChatGPT's virtual clothes try-on feature work?
OpenAI rolled out a global Try On button that uses a selfie, full-body photo, or product screenshot, powered by ChatGPT Images 2.5. Users can also stash looks in a Favorites library, describe a look for an assembled outfit, or match celebrity looks to purchasable items. Instant checkout and agentic product nudges have drawn mixed reactions before, so try-on is the latest shopping bet. Google offered virtual try-on earlier; whether people start fashion journeys inside a chatbot remains open.
What did Microsoft launch with MAI-Transcribe-2-Streaming and MAI voice models?
Microsoft AI released MAI-Transcribe-2-Streaming with first partials in a bit over 100ms, support for 60 languages, continuous language detection, and intro pricing at $0.54 per hour of audio. It also claims the top Artificial Analysis spot for final and partial accuracy. Alongside are MAI-Voice-2.1 at $22 per 1M characters across 23 languages and 26 locales, and Flash at $15 per 1M with about 150ms end-to-end for up to 45 seconds. Access spans Foundry, Playground, Vercel, Azure Voice Live, and OpenRouter.
What is Amazon's Strands Decider 2B and how is it used in AI agents?
Strands Decider 2B is a small, local-friendly decision model open-sourced by AWS Strands Labs. In the Jev mold, it chooses among pre-set options and returns a confidence score instead of a long essay, aiming for a cheaper, more reliable next-step pick in agent workflows. Distinguished engineer Marc Brooker built an early version after seeing TypeSafe's Jev; Amazon cleaned it up for customers on a Qwen3.5-2B base. TypeSafe's CEO warns clones are easy but making them truly sharp is harder.
Why is Wall Street skeptical of Nvidia's chip-backed AI financing plan?
Reuters reports lenders are cooler than Nvidia about treating GPUs like aircraft-grade collateral for a roughly $500 billion financing push with partners including Blackstone, Apollo, and KKR. Nvidia argues top chips can earn for up to a decade; bankers often prefer 3-4 year underwriting, residual-value caution, and stronger guarantees or investment-grade customer cash flows. Pipeline deals may already be tightening terms. CoreWeave's GPU-backed facility leaned on Meta contracts for its rating, which underscores the skeptics' point.
What were the biggest AI news stories on 1 October 2026?
The day's wrap spanned orbit chips, regulation, product launches, and financing. Google's Project Suncatcher put TPUs in low Earth orbit; California's AG subpoenaed OpenAI over cybersecurity risks; and OpenAI cut ties with three safety researchers after alleged confidential leaks. ChatGPT gained a virtual clothes try-on, Microsoft shipped streaming transcription and MAI voice models, Amazon open-sourced Strands Decider 2B, and Wall Street pushed back on Nvidia's GPU-collateral financing thesis.