Three big moves today. OpenAI unveiled its first custom chip, built with Broadcom to run its models more cheaply. Anthropic turned Claude into a tagged Slack teammate that learns your company over time. And a quiet logging bug in OpenAI's Codex was grinding down SSDs before a fix landed. Here is what matters and what to do about it.

OpenAI and Broadcom unveiled Jalapeño, an inference-focused accelerator built to run models like ChatGPT and Codex more efficiently. It went from initial design to manufacturing tape-out in just nine months, one of the fastest ASIC cycles on record, with OpenAI using its own models to help design it. Engineering samples are already running workloads in the lab, with better performance-per-watt claimed versus current state-of-the-art. First production deployment is planned at gigawatt scale alongside Microsoft and other partners before the end of 2026.

Why it matters: Owning the silicon could give OpenAI a durable cost and speed edge on inference, the everyday job of serving answers, while cutting its reliance on third-party GPUs.

What to do:

Source: openai.com

Anthropic launched Claude Tag, which lets anyone summon Claude by typing @Claude in a Slack channel. It reads the thread for context, breaks tasks into steps, and reports back in-thread, while an ambient mode proactively flags stalled work. Admin controls scope what data it can reach per channel. Anthropic says 65% of its own product team's code is now created by its internal version of the tool, which runs on Opus 4.8. It is in beta for Enterprise and Team plans and replaces the older Claude in Slack app, retiring August 3.

Why it matters: Enterprise AI is shifting from standalone chatbots to assistants embedded in the tools teams already use all day.

What to do:

Source: anthropic.com

Garfield AI, the first law firm authorised and regulated by the Solicitors Regulation Authority to operate through AI, helped a freelancer recover £7,000 in unpaid fees. Garfield prepared the pre-action letters, court filings, witness statements, and trial bundle, while a human barrister presented the case in a three-hour hearing at Wandsworth County Court, where a counterclaim was also dismissed. The client paid about £400 in Garfield fees; the losing side used both a solicitor and a barrister.

Why it matters: It shows AI can sharply lower the cost of pursuing small claims that businesses and freelancers often write off, though human advocacy still mattered in court.

What to do:

Source: garfield.law

A logging bug in OpenAI's Codex CLI was generating huge write activity. One developer's main SSD logged about 37 TB of writes in roughly 21 days, extrapolating to around 640 TB a year, near or beyond a typical 1 TB drive's rated lifetime. The cause was a SQLite feedback log running at the most verbose TRACE level and ignoring the usual control for turning it down. The update: on June 22, OpenAI merged fixes that drop the default logging level and cut log writes by roughly 85%.

Why it's interesting: A setting meant for internal debugging shipped to everyone and silently wore down hardware, a reminder that small defaults can carry outsized costs.

Key takeaway: Update Codex; if you ran it heavily for weeks, check your drive's lifetime writes via its SMART data.

Source: theregister.com

A former Google engineer, Justin Poehnelt, says he was fired for building an open-source Google Workspace command-line tool that lets people and AI agents manage Gmail, Drive, Calendar, and Docs from one interface. It hit number one on Hacker News and gathered thousands of GitHub stars in days. The timing stood out: Google announced its own official Workspace CLI at Cloud Next just two days before he was let go. He shared the story publicly this week, about two months after the firing.

Why it's interesting: It captures the tension inside big tech between encouraging grassroots building and controlling what agents might do to an existing business.

Key takeaway: Agent-native interfaces to core SaaS are arriving fast, sometimes faster than the incumbents are ready for.

Source: techcrunch.com

Context: Tied to today's Codex bug, where a recurring problem kept slipping through. Pair every instruction file with a record of known failures so your AI assistant is reminded of corrections on each run.

Prompt: "Create a PITFALLS.md file for [TASK or REPO]. For each known mistake, document three things: the [TRIGGER] that causes it, the [WRONG BEHAVIOR] to avoid, and the [CORRECT BEHAVIOR] to follow instead. Then add a line to the main instructions that references PITFALLS.md and requires checking it before every run."

Do not: Let the file grow into a dumping ground; cap it at the top recurring failures and prune anything no longer relevant.

If / Then: If the same mistake reappears after being logged, then tighten the trigger description and add a concrete example, rather than just restating the rule.

Example: A founder named Priya building an internal coding agent logged the three type errors it kept reintroducing; after referencing PITFALLS.md on each run, those errors stopped recurring across the next two sprints.

Two researchers seen internally as key Gemini contributors, Jonas Adler and Alexander Pritzel, are reported to be joining Anthropic. They are part of a fast run of exits: researcher Noam Shazeer left for OpenAI, and Nobel laureate John Jumper left DeepMind for Anthropic, making four senior departures in roughly six days.

Why it's important: A frontier lab runs on a small number of people who can do the hardest work, and losing several in a week to your closest rivals is the kind of pattern that influences who stays next.

Business takeaway: Pre-IPO equity at Anthropic and OpenAI is reshaping where elite AI talent lands, which shapes which labs ship strongest next.

Source: bloomberg.com

ByteDance unveiled Seedance 2.5, a video model that generates 30-second clips at native 4K from a single prompt, well past the usual 10-to-15-second ceiling, with support for up to 50 multimodal references. It was shown alongside a new language model, image model, and audio model. Two caveats: it is in enterprise beta with a public launch targeted for early July, and the quality claims are vendor-stated, with no independent Seedance 2.5 benchmarks yet.

Why it's important: It underscores how fast Chinese labs are closing the gap on frontier video and multimodal AI, often at lower cost.

Business takeaway: Watch the July launch before committing a video pipeline; if the claims hold, single-pass 30-second clips change the economics of short-form ads.

Source: thenextweb.com

In partnership with EssilorLuxottica, Meta introduced Meta Glasses starting at $299, undercutting its Ray-Ban ($379) and Oakley ($499) lines. They run Meta AI powered by Muse Spark, the first model from Meta Superintelligence Labs, with over 8 hours of battery, live translation, capture, and navigation. The line spans three frame styles and more than 26 color and lens combinations, including a Kylie Jenner collaboration.

Why it's important: Meta is betting low-cost hardware makes glasses the everyday gateway to AI before rivals like Apple ship their own.

Business takeaway: A $299 price point pushes always-on AI wearables toward mainstream adoption, worth tracking if your product could live on a face-worn assistant.

Source: about.fb.com

That's it for today's Daily Pulse. Forward this to one person who wants to stay ahead of AI. See you in the next one. - Nicolas