In partnership with
☀️ TRENDING AI NEWS ⚠️ OpenAI Safety Exodus: David Robinson, who wrote safety reports for every major OpenAI model release, quit this week saying the company's culture is 'broken'. 🏢 California Leads: California became one of the first US states to pass workplace AI protection laws, requiring employers to notify workers before using AI to make job decisions. 🛠️ Meta Opens Muse: Meta open-sourced the code for Muse gadgets, letting any developer build their own AI-powered hardware running the Muse agent. 🚨 Ballot Privacy Risk: Georgia held an emergency meeting after a Princeton researcher showed AI could link anonymous voter ballots back to individual identities. |
Something is clearly wrong inside the world's most prominent AI lab - and this time the person saying so isn't an anonymous source or a disgruntled ex-employee with an axe to grind. He spent years writing the safety reports that accompanied every major OpenAI product launch. That makes today's top story worth paying close attention to.
🤓 AI Trivia
How many spoken words per day did Australians lose on average between 2005 and 2019 - a trend researchers say AI could make even worse?
The answer is hiding near the bottom of today's newsletter... keep scrolling. 👇
| ⚠️ OpenAI's Safety Leader Just Quit - And He's Not Staying Quiet | |
'Not being nearly careful enough' - from someone who was there
David Robinson, who led the writing of safety reports that accompanied every major ChatGPT and model release at OpenAI, has resigned from the company - and he's gone public with his reasons in a piece for The Atlantic. His verdict: AI companies are not 'being nearly careful enough' about developing the technology, and OpenAI's internal culture is 'broken'.
Robinson acknowledged in his own words that he is 'something of a cliche' - another insider who helped build the thing and is now warning about it. But that self-awareness doesn't blunt the message. He had direct, daily visibility into how safety decisions were made at the highest levels of the lab.
A Pattern, Not an Isolated Complaint
Robinson joins a growing list of AI safety voices who have left major labs recently and spoken out. The recurring theme: commercial pressure is consistently winning out over caution. Whether that's a fixable culture problem or a structural one is the question nobody at the top seems keen to answer publicly.
The bottom line
When the person whose literal job was writing your safety reports says the culture is broken, that is not a PR problem - it is a signal that the internal checks at one of the world's most powerful AI labs may not be working the way the public has been led to believe.
Some teams never seem to stop moving. They're on Attio, the agentic CRM.
Every customer signal is captured in one shared context layer, always current and compounding. Agents and workflows build pipeline, chase every buying signal, and move deals forward, an always-on revenue engine running alongside your team.
With Attio, you’ll get:
Leads automatically prioritised and routed to the right rep
Expansion and risk signals caught the moment they land
Follow-ups written in your voice, already there when you arrive
Teams like Parallel, Turbopuffer, and Wordsmith build on Attio. Are you one of them?
| 🏢 California Just Became the Blueprint for AI Worker Protection | |
The first real guardrails on AI in the workplace
As the federal government continues its hands-off approach to AI regulation, California is doing what it often does - moving first. The state has rolled out a set of workplace laws specifically targeting how AI can be used against workers. Employers must notify staff before using AI to make decisions about their jobs, and workers gain new rights to challenge those decisions.
California is, of course, home to most of the companies building the AI that these laws cover. That makes this a particularly pointed move. The state is essentially telling its own backyard tech giants: the AI you are shipping has to meet a bar when it is used against the people who work for you.
Could Ripple Out Across the US
Analysts tracking job automation trends see California's laws as a likely model for other states. With no federal framework in sight, states are filling the vacuum one by one - and whichever framework proves workable in California tends to become the de facto national standard over time.
The bottom line
If you are building AI products that touch hiring, performance reviews, or scheduling, California's new rules are the ones to understand right now - because they are almost certainly coming to your state next.
| 🛠️ Meta Open-Sources Muse So You Can Build Your Own AI Gadgets | |
From your TV to your toaster - Muse wants to be everywhere
Meta has open-sourced the code behind Muse gadgets, giving developers the building blocks to run its new AI agent on their own hardware. The suggested projects range from loading Muse onto a colour E Ink display for reminders, to dropping it on an HDMI stick so it appears on your TV, to building what is essentially a DIY Muse Charm on a small touchscreen device.
The move is part of a broader Meta strategy to get Muse embedded in as many devices as possible - fast. By giving the code away for free, Meta lowers the barrier for hardware makers and hobbyists to add Muse to products that would otherwise require a full SDK deal. It is the classic open-source distribution play: seed the ecosystem, become the default.
The Privacy Trade-Off Nobody Is Talking About Loudly Enough
There is a catch worth flagging. A separate report this week found that Muse builds detailed profiles of your friends and family in order to personalise its responses - a point buried in the fine print of what millions of users have already downloaded. As data privacy concerns around AI agents grow, putting Muse on even more devices will intensify that scrutiny considerably.
The bottom line
If you are a developer curious about building with AI agents on custom hardware, Meta just removed most of the friction - but make sure you understand what Muse is collecting before you ship it to end users.
| ⚠️ Georgia's Emergency Meeting: AI Could Unmask Your Secret Ballot | |
A Princeton researcher just broke ballot anonymity with public data
Georgia held an emergency meeting this week after a Princeton researcher demonstrated something genuinely alarming: publicly available election records, when combined with AI, can be used to link individual voters back to their specific ballots. That is supposed to be impossible. Secret ballots are a foundational guarantee of democratic elections, and this finding suggests the guarantee has a hole in it.
The researcher did not exploit a software bug or hack a system. They used data that is already publicly available - voter registration records combined with ballot sequence information - and applied AI to connect the dots. No breach required.
Why This Hits Differently Than a Typical Security Flaw
Most election security concerns involve protecting systems from external attackers. This one is different: the data is already out there, legally accessible, and the AI is just the analytical layer that makes the connection possible. That makes it much harder to patch quickly. If you have been following our coverage of deepfakes and election integrity, this adds a new and very concrete dimension to how AI threatens the integrity of democratic participation.
The bottom line
This is the kind of AI risk that does not come with a dramatic hack or a villain - just publicly available data, a powerful model, and a privacy guarantee that quietly stopped being true.
| 🔬 Trillium Labs Wants to Do Risky AI Research - Out in the Open | |
While frontier labs lock away dangerous research, one startup is going the other direction
Most frontier AI labs keep their highest-risk research under wraps. Trillium Labs is taking the opposite approach. The startup says it wants to conduct high-stakes work on self-improvement and AI model behaviour completely in the open - publishing findings, methods, and results for the broader research community to scrutinise.
The argument from Trillium's researchers is that locked-away safety research creates a false sense of security. If only a handful of people inside a lab can check the work, the field cannot build genuine consensus about what is actually safe. Transparency, they argue, is itself a safety mechanism.
It is a genuinely interesting counter-bet to the dominant model. And given this week's OpenAI resignation, the timing could not be more pointed. Whether open research on self-improving AI reduces risk or accelerates it is a debate the field has not settled - but Trillium is at least forcing the question into the open.
The bottom line
With AI safety debates happening behind closed doors at most labs, Trillium's bet on radical transparency is either a genuinely important corrective or a dangerous gamble - and the AI safety community will be watching very closely to find out which.
By the way - if any of today's stories have you thinking about spinning up a project or a landing page quickly, 60sec.site is an AI website builder that does exactly what it sounds like. Worth a look if you need something live fast. And for more daily AI coverage, dailyinference.com has the full archive.
| 🌎 Trivia Reveal | |
The answer is 338 words per day! A recent study found that Australians lost an average of 338 spoken words each day between 2005 and 2019 - and researchers warn that increasingly using AI for communication could accelerate that decline further by replacing the conversations we used to have with each other.
| 💬 Quick Question | |
After reading about David Robinson's resignation from OpenAI today - do you think AI safety concerns are being taken seriously enough by the major labs, or does it feel like commercial pressure always wins? Hit reply and tell me what you think. I genuinely read every response and it shapes what we cover next.
That's all for today - a lot of noise in the AI world right now, but the signal is clear: the gap between what labs promise on safety and what insiders say is happening internally keeps widening. We'll keep watching it closely. See you tomorrow!
| 🎁 Share Daily Inference | |
Daily Inference grows by word of mouth. Your personal link is below.