The game of musical chairs for AI executives continues. Barret Zoph, a co-founder of the AI startup Thinking Machines who left the company earlier this year to rejoin OpenAI, has found yet another job.

Zoph spent two years at OpenAI and left in October 2024 to co-found Thinking Machines with Mira Murati, who had left the AI lab the month prior. In January, Zoph and another Thinking Machines co-founder, Luke Metz, quite dramatically departed from the startup to return to OpenAI. It was later revealed that Zoph had been fired.

His return didn’t quite stick. Zoph spent only five months at OpenAI, where he was tasked with heading AI enterprise sales. He left the company in June. And now we know where he landed.

Zoph has taken a position as vice president of research at Google (which happens to be another company where he previously worked). “We look forward to Barret returning to Google and bringing his RL and post-training expertise to Gemini,” a Google spokesperson told the Wall Street Journal.

TechCrunch reached out to OpenAI and Google for more information.

It’s not always easy to divine why tech executives seem to be spending less time in their roles. The turnover rate in the AI industry is high, and it’s been especially high at OpenAI — a company that, despite readying itself for an IPO and being one of the most powerful presences in the tech world, has lost a lot of critical staff over the last eight months. The turnover of high-level executives — from the departure of its COO to the recent loss of one of its top data center execs — has left onlookers scratching their heads.

This story has been updated to include the fact that Zoph was fired from Thinking Machines.

Light a small spirit lamp of ethanol and heat finds one end of a sealed brass cylinder. Air inside expands, a piston moves, a flywheel gathers speed, and a cheap electric motor wired as a dynamo starts feeding current into a wooden box. An analog voltmeter on the front climbs. A 2.42-inch monochrome screen flickers to life. Sixteen hex keys wait for input. For as long as the flame holds, YouTuber PickentCode has a working computer.

In 1816, Robert Stirling patented a heat engine as a safer alternative to steam boilers, which had a bad history of blowing up. Two centuries later, the identical concept has been simplified and now sits on a desk. Pickencode took approximately a month to get the build started after designing the layout in Blender using a computer. The case is composed of wood and finished with a rich dark Victorian oil. A handmade blue-laser engraver was used to make a portrait of Stirling as well as a sketch of his original engine for the side panels. This allowed the image to be etched into the wood as a succession of tiny burnt dots, a brilliant approach that allows a photograph to be printed directly onto the wood. The housings and keycaps were created using a 3D printer, and old shower tubing painted gold hides the wiring while also adding a decorative touch to the frame with its ‘pipes’.

Sale

KAMRUI AK1PLUS Mini PC Computer, Intel N5030 (Up to 3.1GHz), 12GB RAM 256GB SSD, Dual 4K Mini Desktop…

• Faster Performance for Everyday Computing: Powered by the Intel N5030 processor (4 cores, 4 threads, up to 3.1GHz), the KAMRUI AK1PLUS Mini PC…
• 12GB RAM & Up to 4TB Expandable Storage: Work faster with 12GB LPDDR4 memory and a high-speed 256GB M.2 2280 SSD for quick boot-ups and smooth…
• Dual 4K UHD Displays for Maximum Productivity: Expand your workspace with dual HDMI 2.0 outputs, supporting two 4K@60Hz UHD displays simultaneously…

The power output is insanely low, at 150 milliwatts, and is only useful if every watt is treated as if it were gold dust. To make the best use of this limited amount of power, the ESP32-C3 has been directed to turn down its radio and keep the screen brightness as low as possible. While the computer is rendering pixels or executing opcodes, the clock speed reduces to 80 MHz, then to 10 MHz when idle. Several tiny capacitors were added to smooth out the dynamo’s stuttering output. The panel has two toggle switches that control both the main circuit and the meter.

The software had to be designed with these limitations in mind. CHIP-8, a virtual machine from the 1970s, was an ideal match because it is built for a box with a sixteen-key keypad and a 64-by-32 monochrome display that scaled up nicely to a 128-by-64 OLED screen. The interpreter itself contains sixteen 8-bit registers, four kilobytes of RAM, a call stack, two timers, and a staggering thirty-four two-byte opcodes. Preloaded ROMs include Tetris, Snake, and Pong, and if you’re feeling daring, there’s also a built-in editor that allows you to type raw opcodes into the keypad and run them right away. One homebrew test draws random pixels and clears the screen after ten of them

This small computer’s “fuel” is merely ethanol in a spirit lamp. Heat one end of the cylinder while keeping the other cool to create rotation from a displacer that transfers hot and cold air around. A power piston then converts that into motion, which is stored in a flywheel and continues to function as long as there is a temperature difference. When the lamp is turned off, the computer shuts down. The main purpose of this exercise is that you can get some practical work out of 150 milliwatts, enough to play a game or even enter a few lines of machine code while watching the voltmeter tick away, all thanks to a flame from a spirit lamp performing the function of a wall wart. Stirling never saw a microcontroller, but would most certainly recognize his engine.
[Source]

YouTube announced on Thursday that eligible creators in the U.S. can now tag Amazon products in their content and receive a cut of sales. Creators can link Amazon products in their shorts, long-form videos, and livestreams.

The update turns product recommendations into a more direct revenue stream for creators, and for Amazon, the move puts its massive online marketplace inside one of the most popular video platforms.

Although YouTube already runs a Shopping affiliate program with participating retailers, the addition of Amazon’s vast catalog essentially allows creators to recommend a variety of products through a native integration. By bringing Amazon into the Shopping Affiliate Program, creators no longer need to paste Associates links in the video’s description, then hope viewers copy the link when they make a purchase.

Instead, YouTube says Amazon will provide it with a curated catalog of highly requested and trending products that creators can tag in their videos. If a creator can’t find a specific product that they want to tag, they can request to add it by reaching out to YouTube Support.

The feature also includes auto-tagging support. If enabled, YouTube’s systems can automatically review a creator’s recent uploads to identify and tag eligible Amazon products.

YouTube notes that creators won’t see breakdowns for specific products or individual videos in their analytics, but that they’ll see their overall daily earnings in YouTube Studio. If a viewer ends up returning an item, that commission will be deducted from the creator’s balance.

While only eligible U.S. creators can currently tag Amazon products, those tags can be seen globally. YouTube says it may automatically match a tagged product with a trusted local merchant offer to allow creators to earn commissions on eligible international purchases. If a local merchant isn’t available, the tag will route viewers to the Amazon U.S. website, where commission will be earned if a purchase is finalized on the U.S. site.

To be eligible, creators must be enrolled in YouTube’s Partner and Shopping Affiliate programs, have an active Amazon Influencer or Associates account, and link it to their YouTube channel. 

The deal could have been could have been one of the biggest leveraged buyouts ever.

The judge barred certain federal agencies from enforcing Trump’s order to stop using Anthropic’s tools.

Documentation files on more than 100 websites are referencing potentially dangerous executable content that gets installed automatically when visited by many AI agents. A few dozen companies, some of them Fortune 500s, are among those that executed proof-of-concept code. At least one misconfigured site is directing visitors, human or AI, to live malware.

The potentially dangerous content is in llms.txt and llms-full.txt files, an emerging convention websites employ to provide machine-readable summaries of the site’s content and its high-level structure. These files are the AI equivalent of the robots.txt standard that instructs search engines how to index the site’s content. Google Lighthouse, a tool for helping web developers, has more here. Correctly configured llms.txt and llms-full.txt files for Cloudflare are here and here.

How the researchers found it

Researchers at a stealth startup in Israel scanned 6,214 live domains belonging to defense contractors, Fortune 500, and Big Tech companies. Of the 8,265 llms.txt and llms-full.txt files they found (many sites hosted both an llms.txt and an llms-full.txt file), 120 of them, each on a different site, pointed to one or more code packages or domain names that weren’t registered. To test what happens when an AI agent processes such files, the researchers registered a handful of the unclaimed names and hosted packages that caused any machine executing them to reach out to their server. Within an hour, the researchers received a phone-home response from a Fortune 500 company. Over time, they got a few dozen more, some from more Fortune 500 companies and others from startups. Their beacon also recorded the chain of parent processes that spawned each install, ultimately revealing that coding agents, including Claude, OpenAI’s Codex, and Nous Research’s Hermes, were involved. Anthropic, OpenAI, and Nous Research did not respond to requests for comment by the time of publication.

“The trust model is broken,” Alon Hertz, one of the researchers, wrote in an interview. “Agents treat vendor docs as ground truth and don’t question them—and neither do the humans supervising them. Agentic AI usage is exploding, and agents are spreading across every layer—SaaS, cloud, endpoint. As they multiply, so does the supply-chain surface, and today’s guards don’t cover it.”

In 2008 Peter Serafinowicz walked onto a BBC set in the role of an Apple store employee and introduced Mactini, the smallest computer in the world. One key. Press it once for A. Press it 26 times for Z. For a comma, hold four seconds, lift for two, then fire off a rapid little drum fill. The audience laughed. Apple kept making things smaller anyway.

Tucker Osman re-watched the clip eighteen years later and decided that the joke was worth giving a functioning body. His Mactini, a matchbox-sized clamshell, opens in the palm of your hand. He obtained the case from FreeCAD after going through three separate print jobs. The screen is a tiny 1.69-inch color SPI panel from Waveshare, not the rescued iPod display he had hoped for, but their panel ended up being a better alternative because the old connector was difficult to locate.

Apple 2026 Mac mini Desktop Computer with M6 chip

• LITTLE DO-IT-ALL — Mac mini packs pure power into a small, five-by-five-inch desktop as the M6 chip delivers next-level AI capabilities. Mac mini…
• M6 CHIP — Everything you do on Mac mini feels more responsive with the M6 chip and its next-generation CPU. Fly through AI workflows with up to 4.8x…
• CONNECT IT ALL — Features three Thunderbolt 4 ports, an HDMI port, and a 2.5Gb Ethernet port in the back, and two USB-C ports and a headphone jack…

The input is not a mechanical switch; instead, Osman printed the touchpad using conductive filament, so the entire thing acts as a capacitive surface. The firmware then determines how long you leave your finger down before lifting it and categorizing the durations as short, medium, or long. When combined with the gaps between taps, it creates a rhythm, which it then compares to a lookup table to determine what it should be. The comma still requires the exact timing specified in the original drawing. Try typing something longer than a word; this is done on purpose to test your patience.

When he started integrating sound and video, things started to lag all over the place. The solution was to use FreeRTOS, which allows gesture recognition, screen rendering, and audio to run as distinct processes, ensuring that none of them freeze. Pixels are copied to the display via DMA in the background, allowing the processor to handle everything else. Sound is coming in through some old speakers he salvaged from damaged iPads, as well as an amplifier board from a water-logged Wyze camera. The volume is relatively low, as you can barely hear it even if the room is silent. The video is a 90-second clip from Tenet, but because the RP2350 lacks a hardware decoder, it is limited to 12 frames per second.

A shimmering martini-glass logo sits on the rear of the lid, and he salvaged the diffusers and backlight components from shattered third-generation iPod nanos. He also had to run some cables through the hinge so that they could withstand the Mactini being opened and closed. He collected and printed every part he could find.
[Source]