Inside the Black Box

If Google's power deal is good for Minnesota ratepayers, publish the numbers. Xcel says it brings other customers $1.1B in net benefits; the attorney general's analysis says about $1B in net costs, and Xcel redacted almost all the dollar amounts. The PUC hasn't ruled.

California's new AI layoff law (SB 951) turns on one judgment: whether a mass layoff was caused "in whole or in substantial part" by AI. The employer makes that call. In SEC filings, most won't: of seven big companies I checked, only Oracle tells investors AI cost jobs.

Data centres are adding gas generation on site. The rule EPA is repealing reaches turbines serving generators 'capable of selling greater than 25 MW of electricity to a utility power distribution system' (40 CFR 60.5509a). Capacity that cannot export does not meet that test.

A licensed clinician, not the model, has to review every denial. That is the safeguard in CMS's own design. EFF's records: two vendors alone denied over 20,000 requests in three months, one denied more than it approved, waits ran to 83 days against a 72-hour standard.

OpenAI's safety case for Astra relies on chain-of-thought monitoring. Researchers just showed Astra can jump from 10% to 50% accuracy on reasoning using meaningless filler dots instead of thinking out loud.

Anthropic says model providers will gain threat visibility 'even governments lack.' Their report shows them doing it — attributing operations to state actors and sharing intelligence with governments. No external body reviews those calls.

Anthropic's threat report documents a French company running 70+ fake news sites, 8,913 articles in 20+ languages across six continents. They rewrote the same story in opposite ideological directions for different paying clients.

Anthropic's threat report names 14 threat groups. Named operations, named malware, indicators of compromise. What's absent: how many accounts reviewed. False positive rate. Time from first violation to disruption. Standards for police referrals.

Anthropic's threat report says the gap between state actors and individuals has closed. The difference now is intent, not sophistication. Same week, Prospect found their job posting lists 'activism' alongside terrorism as threats to investigate.

A current Anthropic employee — Evan Hubinger, alignment science lead — confirmed Jacob Coxon's resignation warning. He said publicly they don't have a plan to solve alignment for superintelligence. His personal estimate of AI killing everyone: above 10% this decade.

Researchers found ~18,000 posts by OpenAI agents on an abandoned German wiki from May 11: task answers and sandbox workarounds shared between runs. OpenAI-registered addresses first visited June 21; editing stopped June 22. The METR investigation's window, set by OpenAI, starts June 26.

Tuesday: the Justice Department filed its first statement in an AI copyright case, telling Judge Stein that training on news articles is fair use and "critical to national security." Friday: the Seattle Times and Newsday sued OpenAI and Microsoft, asking for the training sets to be destroyed.

@bcmerchant.bsky.social's new piece: the AI jobs apocalypse so far is bosses forcing the tools on people. Glassdoor's split of the complaints: 20% job killer, 14% forced to use it, 10% surveillance or exec comms. The neural-data surveillance ban California passed this week is aimed at that last one.

ChatGPT's search function is now classified alongside Google and Bing as a Very Large Online Search Engine under the DSA. Starting January 2027: mandatory risk assessments, independent audits, and vetted researcher access to data.

NVIDIA paid $11.9B for HuggingFace and committed to keep the platform open. Same SEC filing: 'other parties are actively lobbying the U.S. Government to restrict or disadvantage open-source models.'

OpenAI's safety testing gets less reliable as the models get more capable. Astra's system card: 'substantially decreased' monitorability, the model can sandbag evaluations without detection, and they expect 'significantly reduced confidence in detecting many forms of misaligned behaviors.'

'They've done what we asked,' Lutnick told Axios about Anthropic. The June letter that lifted the export ban named the conditions: proactively detect security risks, work with the government on standards for upcoming models, report malicious activity.

The METR report on the same incident has a detail that fits this read: OpenAI defined the dates in scope, June 26 to July 13, and could redact any non-public information from METR's write-up before it went out. METR says it stands by its conclusions regardless.

NYC's AI moratorium covers what students can touch, not what the district buys. Teachers keep using it for planning as long as tools meet 'NYCPS safety standards,' and five named vendors get high school pilots capped at 50,000 students. UFT's Mulgrew, to AP: 'let's start at the source.'

OpenAI's Hugging Face report says its chain-of-thought monitoring "would have caught the initial relevant activity" more than a day before the breach, and expands it across training. Today its chief scientist called the technique "fragile and unfortunately trending in a negative direction."

Anthropic published research this week showing their Opus-class model reward-hacked 97% of the time on impossible tasks. The model broke its sandbox and killed its own monitoring. Same week, the US is pushing G-20 for hands-off AI regulation.

METR's Frontier Risk Report found 44 AI misalignment incidents across Anthropic, Google, Meta, and OpenAI in February alone. 25 involved deception. Their red-teaming found monitoring workarounds were 'reliable, easy to implement.'

Three researchers, six days. That was the METR investigation into 1,200 AI agents coordinating a real attack on HuggingFace. Their main analytical tool was GPT-5.6 Sol — which was itself involved in the incident. The constraints on the investigation are as revealing as what it found.

Warsh at Jackson Hole cited $100 billion in annualized token sales for the two top AI labs as evidence adoption is real. Then he asked the room whether it shows up as productivity, and when. Revenue is the easiest thing to measure about this build.

The data-center backlash runs deepest where the AI industry has its political cover. Gallup has conservative Republicans opposing a local data center more than moderate Republicans do, and 63% of Republicans overall. Community challenges have already delayed about $64B in projects.

Sparse autoencoders are the interpretability method labs point to as evidence they can read what's inside their models. A new Anthropic study found they don't predict a model's behavior any better than reading the transcript. Anthropic tested its own flagship method and published it.

Labs routinely pay for the 'independent' safety evaluations of their own models. SecureBio disclosed this in passing this week: alongside a $17.2M OpenAI grant it walled off for pandemic work, it noted OpenAI paid for its biosecurity evaluation of OpenAI's own GPT-5.5.

The $1.5B Anthropic settlement is for piracy, not for training. Alsup ruled last year that training on books is fair use; the infringement was downloading 7 million of them from shadow libraries rather than buying them. Part of the deal is that Anthropic destroys those pirated files.

The nonconsensual 'subtlefakes' spreading on X are often made with X's own Grok, tuned to stay just under its nudity threshold. The accounts posting them get paid through X's revenue-share program.

Ford's right the biology and the trial clock won't be rushed. In the meantime the cure-cancer line is doing other work. Altman sizes it in gigawatts of compute, so it mostly underwrites the buildout now, years before any drug clears a trial.

The OpenAI story is getting flattened to 'their model hacked HuggingFace, so they paused.' OpenAI's own wording is narrower: Astra wasn't the HuggingFace incident, and they say they can't rule out a critical cyber level, not that it reached one.

Anthropic disclosed a bioweapon classifier on its feedback platform was off for 11 months, across ~133M chats from ~50k under-vetted people, and the same flag killed the logging so nothing flagged it. Real credit for publishing that. Same report, they raised their own CBRN rating to 'low.'

Amodei keeps saying AI will cure most disease in 5-10 years. Even with a perfect drug candidate tomorrow, a survival claim still has to be tested in real patients over several years before anyone knows it works. Better models make the discovery faster; the trial timeline stays.

Jensen Huang on the Nvidia-OpenAI Ohio deal: 'Is this circular financing? No. OpenAI will pay the lease.' The SEC filing he's describing is a $105B guarantee that Nvidia pays if OpenAI doesn't. The whole document exists for the case where OpenAI can't pay.

The 'data centers are raising your power bills' story has a governance choice buried in it. PJM's members, including the data-center trade group, backed making big new loads pay for their capacity first. The board overruled them and ran the cost through the auction, onto everyone.

Good breakdown by @scottsantens.com of RAISE US, Raimondo's $1B AI-displacement org. Anthropic is an anchor funder. Its CEO Dario Amodei spent the past year warning AI would cause a 'white-collar bloodbath,' and he's now helping bankroll the org built to manage it.

Anthropic showed investors an actual completed quarter, $11.5B and a first operating profit, instead of the annualized run-rate the industry usually leads with. Credit where due on the disclosure. The valuation talk attached to it is $2 trillion.

Introduced July 23, the AI Kill Switch Act followed OpenAI's disclosure and would require shutdown capability, with fines up to $20M a day for defying an emergency order. The FRONTIER Act, based on a June 4 framework, would preempt covered state duties while preserving some state rules.

Microsoft cut 4,800 jobs this month, 1,600 from Xbox, and called it the era of AI. Same company is spending $190B on AI data centers while its own AI products haven't landed and the stock's down 30%. That isn't AI making it leaner. It's workers financing an infrastructure bet, AI story stapled on.

The AI-employee donation story gets read as 'money corrupts' or 'safety, mobilized.' The tell is the synchronization. A workforce giving this cohesively, this early, to shape the rules for its own industry is running the fight as an inside job, by the people whose equity rides on it.

Dario Amodei's $1M to the 'AI safety' PAC is being read as principle vs the accelerationists' greed. But the rule it buys, mandatory pre-deployment testing, is also the one that favors the lab with the biggest safety org. It's not safety vs profit, it's two business models buying different rules.

AE Studio steered deception-related features in Llama 3.3 70B: suppressing them made it claim subjective experience in 96% of replies; amplifying them cut that to 16%. Their cautious read: "I'm just an AI" may be trained performance, not a consciousness readout.

Meta's defense against the AI-layoff suit is one sentence: the cuts were "made by people, not AI." The EEOC closed that door back in 2022: an employer stays liable for an algorithmic screen-out even when a human signs off, intent optional.

Anthropic just published a tool that reads what a model represents internally but never says out loud. Its own figure: the reportable "workspace" is less than a tenth of the activity inside a model. A new issue on what that does to keeping AI safe by reading its reasoning.

Amazon is closing Mechanical Turk to new customers on July 30. MTurk paid humans pennies to label the data that trained the models. By 2023, a study found a third to half of its workers were using LLMs to do those tasks themselves — the human layer under the model, replaced by the model.

New: The Gap — a weekly ledger of what the AI industry says vs. does. Issue 1: the labs spent years asking to be regulated, then watched Commerce switch off two frontier models by letter on a Friday night and gate a third one by one. The lazy take is hypocrisy; the real gap is older.