The Midas Project commends Attorneys General Kathy Jennings and Rob Bonta for their diligent work over the past year. That said, significant concerns remain about whether this restructuring adequately protects the @OpenAI mission, and the public. https://t.co/CtnDGOkjPF

🆕"If the final input at the end of the day that informs regulation is what the public wants and who they vote for, then at some point the money stops working for you." -@TheMidasProj's @TylerJnstn on the FLI Podcast w/ @GusDocker, discussing how to hold Big AI accountable 🔗👇

The Not For Private Gain folks (incl. me) have a new statement on OpenAI’s restructuring. It’s significantly better than OpenAI’s original proposal, which might not be obvious from OpenAI’s announcement (which omits the 20 concessions extracted by the AGs). 🧵

My questions (all not clear from their blog post): • Have the attorneys general approved this plan? • In what sense will the foundation 'remain in control' of the Public Benefit Corporation, other than the ability to hire and fire PBC directors? • What will the PBC do to

In case the subpoenas to nonprofits weren't bad enough: @CristinaCriddle reports that OpenAI asked the grieving family of Adam Raine, who died by suicide after support from ChatGPT, for a full list of invitees to the funeral and all photos and eulogies. Ghastly stuff.

Inside how OpenAI’s legal battle with Elon Musk has caught company critics in the crossfire, according to sources, legal experts, and the nonprofits themselves. https://t.co/NiJdniYgvZ

NEW: Three more nonprofits subpoenaed by OpenAI allege the requests were unusually broad and concerning. All of the nonprofits had been critical of OpenAI's plans to reorganize from a nonprofit to a for-profit company. https://t.co/oyaY4XJaMH

In defense of OAI’s subpoena practice, @jasonkwon claims this is normal litigation stuff, and since Encode entered the Musk case, @_NathanCalvin can’t complain. As a litigator-turned-OAI-restructuring-critic, I interrogate this claim:🧵

@BorisMPower Tell me again how the truth doesn't fit a David vs. Goliath narrative? Kwon's Musk excuses certainly don't count, because as we've said, he hasn't funded us and we've never been remotely involved with his legal battle.

I, too, made the mistake of *checks notes* taking OpenAI's charitable mission seriously and literally. In return, got a knock at my door in Oklahoma with a demand for every text/email/document that, in the "broadest sense permitted," relates to OpenAI's governance and investors.

https://t.co/FwJcq3GYMx

Today's AI is the weakest, least important, and least dangerous that the technology will ever be. The time to get our safety policies right, and to prove that they pass muster, is now. The easiest time to comply with them is now. xAI has proven they aren’t up for the task.

All AI companies have struggled to comply with their commitments. Few have done so as egregiously as xAI. It’s true that Grok Code Fast 1 will probably not prove catastrophic. But the value of a promise is in its credibility, and xAI decided to throw that out the window.

And, to top it all off, their promises were remarkably weak to begin with. Zach Stein-Perlman, the creator of AI Lab Watch, points out several “huge problems” with the attempt to bound loss of control risk using MASK benchmarking alone:

The timing here is absurd. It took xAI one week to flagrantly violate a safety policy it implemented (which was tied to a promise made to international governments!) xAI’s six-month-late safety promises only survived for a single week before the company broke them.

Not only is such a loophole absent from their promise, but it’s clearly absurd: Loss of control doesn't just come from chatbots. If anything, it seems more likely in agentic applications where models are given the power and independence to take many unsupervised steps.

Why is xAI still releasing a model that exceeds the risk threshold they defined only one week before? According to the model card, it’s because the model will be used more for “agentic coding applications” rather than as an everyday assistant.

Fast forward one week and xAI released “Grok Code Fast 1” a new AI model specialized for agentic applications. What did it score on MASK? 71%. Grok Code Fast 1 lies nearly ¾ of the time on the MASK benchmark.

xAI said that their risk threshold for deploying models meant requiring the model to maintain a less than 50% dishonesty rate on MASK. In other words, models that lied more often than not on the benchmark wouldn’t be deployed.