Texas wants an outside evaluator approving every ChatGPT model before Texans touch it

Good morning 👋 Texas just asked a judge to put a permitting gate in front of every future ChatGPT release, not just the usual slap on the wrist.

In today's issue:

• 🔭 Texas wants an outside evaluator approving every ChatGPT model before Texans touch it
• 🧠 Models & Releases: Rescale skips the orchestration slog with AWS's AI-DLC
• 🔬 Research Highlights: an inbox agent's real error rate, greener code that burns more watts, and who the free tier is actually for
• 🛠️ Try This: give your own model a sign-off before anyone forces you to
• 🔗 Quick Links: OpenAI's $30B round comes in lighter than advertised

🔭 THE ONE THING

⚖️ Texas wants an outside evaluator approving every ChatGPT model before Texans touch it

Ken Paxton filed a 67-page petition against OpenAI Global LLC and Sam Altman personally in Collin County district court on Oct. 23, leaning on the state's deceptive-trade-practices law and the SCOPE Act. Banning minors and forcing real age checks are in there too (the free tier runs none today, the paid tier asks and never verifies), but the injunction that should worry OpenAI's lawyers is narrower: no new model reaches Texas without sign-off from an outside evaluator first. Texas is at least the tenth state to sue OpenAI this year, after Florida in June and eight more in mid-October, and every prior complaint reads like a liability case. This one reads like a permitting regime. Grant it, and OpenAI's release calendar starts running through a Collin County courtroom, with every other state AG watching for the template.

🧠 MODELS & RELEASES

• 🤖 Rescale stood up a working test-triage agent in weeks, not quarters, by running AWS's new AI-DLC framework with Strands Agents and Bedrock instead of hand-rolling the orchestration. Builder read: the framework, not the model, is the bottleneck most teams are actually stuck on. AWS blog

🔬 RESEARCH HIGHLIGHTS

• Foyer's inbox benchmark turned error-checking into the whole study. A self-hosted email agent, tested against 739 real messages with human labels, found the production keyword rules scored 28% accuracy against a 58% majority-class baseline: the field it shows most prominently is the one it gets most wrong. Plain logistic regression trained on just 200 labels beat every LLM on raw accuracy, few-shot prompting included, but that was the wrong scoreboard: by accuracy it was best, yet it ignored 18% of emails that actually needed action versus 4% for the strongest LLM. A cost-tuned threshold on its noise score halved that miss rate and left it the cheapest system in the lineup. arxiv.org/abs/2610.32955v1

• Faster code, dirtier watts. Researchers ran 11 open models through 1,000 EffiBench problems and measured what their "optimized" code costs to actually execute. 42 of 44 configurations burned more energy than the human-written baseline, overhead running as high as 886%. Mixture-of-experts models were the exception, cheaper to run without losing accuracy. If your team ships LLM-refactored code assuming faster means greener, this is the paper that says check the power meter, not just the clock. arxiv.org/abs/2610.32891v1

• Who pays for your free tier. A pricing model of ad-supported LLM platforms finds the ads aren't really there to monetize free users. Degrading the middle, ad-heavy plan makes the paid tier look better by comparison, letting the platform charge premium users more without touching the premium product at all. In the model, an intermediate ad cap, not a ban, is what maximizes consumer surplus. Worth remembering next time a "generous" free tier gets a little worse. arxiv.org/abs/2610.32945v1

🛠️ TRY THIS

Give your own model a sign-off before anyone forces you to

1. Open Hugging Face's torch-neuronx model support Space and search for whatever model you're about to push to production.

2. Match it to your stack first. Set the Neuron compiler version and Transformers version you actually run, or the pass/fail flag is meaningless.

3. Read past pass/fail. The real numbers are cores needed, wall-clock time, peak memory. That's your migration cost, not a vendor slide.

4. Pull the Space's run.py and NEFF cache and reproduce the result yourself before you move anything off GPU. Coverage today: text-generation and image-text-to-text models, capped at 16 cores.

Prompt: Given this Hugging Face model ID, check the aws-neuron/torch-neuronx-model-support dataset for pass/fail status, Trainium cores required, and peak memory, then tell me if it's worth migrating off GPU.

🔗 QUICK LINKS

• OpenAI is wrapping a $30B round led by UAE investors and Coatue at a $1.25T valuation, down from the $1.4T it was shopping just weeks ago. CNBC

See you tomorrow.

Pradeep Perugu