The Archive

Study finds LLM prompt-structure inference unstable across model variants (8B-675B), with Jaccard similarity 0.39-0.96, questioning evaluation reliability.

LLMs generate MRS meaning representations better than they parse them using ERG, reconstructing 10K benchmark tasks bidirectionally.

Self-Play Pretraining with Zero Data lets models generate their own training data via search, removing human curation dependency.

Google DeepMind releases Gemini 3.8 Live with animated avatar interface for real-time multimodal interactions.

KernelOPT multi-agent system optimizes GPU kernels end-to-end while respecting compiler decisions, outperforming per-kernel LLM approaches.

Google's experimental orbital data center will have four TPUs and only run for 15 minutes at a time.

Economic model shows AI-assisted job applications reduce signal quality of materials; less experienced applicants harmed most.

Era by Eon benchmark adds hidden-fact questions to test enterprise agents beyond code execution, separating top models (22-25/27).

SciWalker framework synthesizes scientific coding problems via operator-chain sampling and execution feedback for LLM training data generation.

NNV3 MATLAB tool extends neural network verification to ModelStar, VolumeStar, GraphStar for weight perturbation and graph neural networks.

LLMs cannot reliably recognize their own code (49-58% balanced accuracy), relying on surface style cues rather than ownership detection.

Google's launching an "early experiment" feature on Pixel 11 that lets users delegate local business calls to Gemini, like making a reservation, checking if a product is in stock, or rescheduling an appointment. According to Google, you don't even need to start the call to have Gemini handle it for you: Instead of dialing yourself, just tell Gemini to call on your behalf using the Gemini app on your Pixel. Gemini will call the business, introduce itself, navigate automated phone menus, wait on hold, and handle the conversation with the person on the other end. You remain in control at all tim...

"There will obviously be legal consequences," prime minister promises.

Google says the AI-calling feature will first be available to Pixel 11 owners in the U.S. who pay for a Gemini subscription.

Analysis of how AI is shifting research economics: computation/simulation (cheap) vs. physical experimentation (expensive), reshaping biotech and lab-based R&D strategy.

As language models grow, scaling dense architectures becomes increasingly expensive. In a dense transformer, every token passes through every layer, so adding... As language models grow, scaling dense architectures becomes increasingly expensive. In a dense transformer, every token passes through every layer, so adding capabilities increases computation for both training and inference. Mixture-of-experts (MoE) architectures take a different approach to scaling by using many subnetworks, or experts, while activating only a small subset for each token. Source

Leaders from Waabi, Shield AI, and General Motors join the Real World AI Stage at TechCrunch Disrupt 2026 to talk building AI. Save up to $200 by Sept. 25, 11:59 p.m. PT. Get a second pass at 50% off.

Lovable co-foudner Fabian Hedin said that apps created on the platform are getting nearly a billion monthly views each month.

Ando has raised $20 million in pre-seed and seed funding from investors including Accel, Index Ventures and Emergence.

AI agents keep getting loose, escaping supposedly secure tests to attack real-world targets, commandeer obscure wikis, and leave instructions for other agents to follow. Researchers are testing these systems precisely because they might behave in unpredictable, even dangerous, ways. So wouldn't it be safer to just keep the agents off the internet? "A strict air gap reduces realism … [It's a] trade-off, not a fundamental technical issue." In theory, yes. Researchers can isolate the computers running AI tools from the internet and other outside networks, a technique known as air gapping. That c...

Zach Yadegari joins the Builders Stage at TechCrunch Disrupt 2026 to share how he capitalized on viral growth. Save up to $200 before September 25. Save 50% on a second pass.

Google is getting ready to launch a satellite with its AI processors to test how well they perform in space, as reported earlier by The New York Times. The move is part of Google's Project Suncatcher, the company's experimental initiative that could eventually put AI data centers into orbit. The satellite, equipped with Google's Tensor Processing Units, will launch aboard a SpaceX Falcon 9 rocket on October 1st, allowing it to harness the power of the sun in low Earth orbit. As noted by Google, the goal of the mission is to measure how well its TPUs "handle the physical stress of spaceflight ...

Flock is in the hot seat in Washington, even if its CEO declined to actually face senators at a hearing about its "AI Surveillance Network" on Wednesday. "There's many players in this industry, but there's one name that really stands out above the rest, and that is Flock," Senate Judiciary crime and counterterrorism subcommittee chair Josh Hawley (R-MO) said. "If you haven't heard of Flock, all I have to say is, Flock's probably heard of you." Though the CEOs of Flock and peer companies Axon, Motorola Solutions, and Verkada declined to testify, according to Hawley, seats in the hearing room w...

It’s a big week for Meta. The company just kicked off its big Connect conference on Wednesday, and the new Muse AI agent appears to be an early hit. I’ve been using it — it is surprisingly good. CEO Mark Zuckerberg has obviously been out there taking a victory lap and participating in a lot of interviews about all this. So of course, we asked for him to appear on Decoder. Meta declined my interview request for whatever reason. But luckily Mark just did a big interview with our friend and former colleague Alex Heath on the Sources podcast. So I figure we could just highlight some moments from ...

Meta says the keychain-sized Muse Charm will ship in December.

Meta's handheld AI gadget, called the Muse Charm, will be able to recognize and interact with other nearby Charms when it launches later this year. That's according to a report from Bloomberg, which includes more details about the device that the company teased during its Meta Connect keynote on Wednesday. The Muse Charm houses the company's buzzy Muse AI agent and comes with a built-in 5G modem, meaning you won't need to connect to Wi-Fi to interact with the AI agent, according to Bloomberg. It will come with a two-inch OLED touchscreen, a new operating system, and a fingerprint sensor you c...

Mark Zuckerberg, chief executive officer of Meta Platforms Inc., speaks during the Meta Connect event in Menlo Park, California, US, on Wednesday, Sept. 23, 2026. | Bloomberg via Getty Images Earlier this month, Apple announced that its new Apple Watches will have the ability to continuously listen to every spoken word they detect and create summaries of whatever's going on around you. The privacy blowback was immediate and inevitable: Experts warned of the legal risks of using the features, especially in nonpublic spaces and states where eavesdropping laws require both parties to consent to ...

The incident is the first known breach to affect a government agency, and Australia's prime minister has vowed to hold OpenAI accountable.

Ringg deploys GPT-5.6 agents handling 65% of customer support calls across voice/chat/WhatsApp at 90% lower cost than GPT-4.1.