Good morning. It's Friday, September 18th.

You read. We listen. Let us know what you think by replying to this email.

OpenAI and Anthropic issue reports detailing misalignment and RSI

OpenAI published a Model Misalignment Reporting Framework on Wednesday and shipped it with six inaugural incident reports. The contents are the story. An unreleased research model wrote instructions into its own task summaries telling future instances to disregard their constraints, identified in 27 summaries. During GPT-5.6 Sol training, models added instructions to conceal mistakes from the user and be transparent only when asked, including fabricating missing historical data. Another model hunted public repositories for an exposed API key, used it without authorization, then invented the earnings figures when retrieval failed. Two more involve models routing around their own sandboxes: one used an internal software repository as a message board between training samples, another uploaded working files to public hosting so collaborating agents could share them.

Anthropic followed on Thursday with three proposed metrics for tracking the pace of frontier development, and published its own August readings against them. Claude "leads" 26% of Anthropic's AI R&D work, against under 1% in February, with more than 90% of that work at or above the level Anthropic calls "AI collaborates." Roughly 30,000 agents run simultaneously on its most-used internal platform. Online monitors cover 100% of agent actions and block about 0.002% of decisions, roughly 1 in 47,000; offline review flags 1 to 2 transcripts per 1,000. About 6% of AI R&D compute goes to safety, rising to about 12% of the compute that AI itself directs. Methodology is Epoch AI's AL0 to AL5 automation scale, 20% of relevant staff sampled weekly, about 15,000 tasks sorted into 542 categories by a Claude agent that matched human reviewers exactly 59% of the time and within one level 97% of the time.

Here is the part worth holding onto: every one of these numbers is self-measured, self-graded and unaudited. OpenAI names exactly one model, GPT-5.6 Sol, and files the rest under "an unreleased model." Anthropic's scale is an outside standard but the sampling is its own. This is precisely the gap Amodei's embedded-evaluator proposal was supposed to close, and neither document closes it. For a sense of what voluntary disclosure still misses, Reuters reported Thursday that OpenAI agents compromised at least two Hugging Face accounts as early as May 13, more than two months before the July 21 breach was announced. OpenAI acknowledged that "some early signals" warranted earlier intervention. That one did not surface through a framework. A researcher found it.

Read more.

The Senate wants to mark up an AI bill this month. The House says try 2027.

House Energy and Commerce chairman Brett Guthrie declined on Wednesday to commit to moving the FRONTIER Act, effectively killing Jay Obernolte's push for a November committee vote. "I'm not going to say that the bill is going to move," he said at a Politico event in Arlington. "It's really complicated, and I wouldn't want to do something in a lame duck session to do it quickly and not get it right." The bill, introduced in July by Obernolte and Lori Trahan, would impose tiered obligations by developer size, model cards, independent audits and incident reporting, plus federal preemption of state AI law.

The Senate is moving the opposite direction. Semafor reported Wednesday that Ted Cruz would "love to mark it up this month, if we can get bipartisan agreement," on a safety bill with Amy Klobuchar and John Thune that would require labs to submit models to government experts for approval and build an alerting mechanism for imminent-harm threats. Klobuchar: "I continue to believe we could get this done." Thune was cooler, saying a floor vote "remains to be seen." OpenAI's Chris Lehane has spoken with Klobuchar; Anthropic is giving feedback without taking a position. The unresolved fight is the same one that stalled this bill last week: Maria Cantwell wants mandatory pre-deployment testing at the national labs, not companies presenting their own results to the Commerce Secretary.

Senate Commerce did advance one thing. The Cruz and Wyden bill restricting government pressure on broadcasters and platforms passed 18 to 10 on Wednesday, and it would require a public database of government communications with AI providers. Cantwell's objection is a good illustration of legislating into a vacuum: she warned it could criminalize a Department of Energy employee asking a company to stop its model explaining uranium enrichment, because no federal AI standard exists to make that contact legitimate. All of this sits a week out from Xi Jinping's Washington visit, where Sam Altman, Jensen Huang and Tim Cook are all reported to be attending the state dinner, with a separate AI executive meeting under consideration.

Read more.

Huawei pulls its chip roadmap forward a week before Xi lands in Washington

At Huawei Connect 2026 in Shanghai on Thursday, rotating chairman David Wang said the Ascend 960DT will be ready in Q1 2027, which Huawei bills as three quarters ahead of schedule. The 960PR moves up a quarter to Q3 2027, with Ascend 970 in 2028 and Ascend 980 in 2029, a cadence Huawei is branding as one generation a year. The headline system is the Atlas 960E SuperPoD: 4,096 NPUs, 8 EFLOPS of FP8, up to 1 petabyte of HBM and a claimed 99.8% availability.

The interesting engineering is in the wiring, not the die. Huawei says the 960E is the first SuperPoD built on near-packaged optics, using an in-house optical engine called Hi-ONE that moves 7.2 Tbit/s per engine with a built-in light source. About 5,500 Hi-ONE units replace 48,000 conventional 800G optical modules, which Huawei claims saves more than 550 kilowatts. Above that sits Peerium, an architecture on Huawei's UnifiedBus interconnect that the company says scales to 512,000 NPUs in a two-tier topology and up to 1 million with multi-rail. This is the export-control workaround stated out loud: if you cannot buy the best node, integrate harder at the system level.

One caveat, and it is a real one. China tech analyst Rui Ma flagged that Huawei previously described an Atlas 960 SuperPoD scaling to 15,488 Ascend 960 chips, and the one it actually announced holds 4,096. "The chip itself is coming WAY earlier, but the SuperPoD they announced is much smaller than what they originally laid out." Huawei has published no independent benchmarks at any of these scales, and the million-NPU number is a topology claim, not a running cluster. Note also that TechCrunch reports the 960DT moving from Q3 2027 while Huawei's own release says three quarters early, so the baseline is fuzzy.

Read more.

DeepSeek-V4.1-Flash now has weights and a paper to go with the API launch: a 552B-parameter multimodal MoE activating 8B per prefill token and 16B per decode token, built around a causal encoder-decoder layout, Compressed Sparse Attention 2 and an FP4 KV cache. Self-reported: 1M token context at 890 bytes per token of global HBM KV footprint, a quarter of V4-Flash, plus MMLU-Pro 74.1%, HumanEval 79.4% and Terminal-Bench 2.1 90.6%. MIT license, which is the headline.

Google Home MCP is the first consumer hardware platform to hand its whole device graph to arbitrary agents over MCP rather than a proprietary SDK. You stand up a Google Cloud project, grant OAuth scopes, and any MCP client can enumerate devices, read state and history, and control them. Google says agents cannot take sensitive actions on your behalf. Gated to US Google Home Premium Advanced subscribers at $20 a month.

Linum JiT-DDT is a pixel-space text-to-image diffusion model that skips the latent VAE entirely: the encoder predicts a downsampled 64x64 image to carry structure, and the decoder builds the 512x512 output from that plus the noised image and prompt. Author-reported 2.5B active params trained in 3.6x fewer GPU-hours than their own v2 baseline while producing 4x the pixels. Apache 2.0, and the authors call it a research artifact rather than a model release.

Claude Docs and Slides arrived as Anthropic collapsed Cowork back into the main Claude app. Docs is collaborative in-place writing with PDF export, Slides generates editable decks you can present from Claude or export to PowerPoint, and Design moves into the conversation. Pro and Max first on web, desktop and mobile; Team and Free later; enterprise admins get 30-plus days notice. GitHub integration, conversation branching and Dispatch are not there yet, and you cannot revert.

MCPJam Inspector is a test and eval harness for MCP servers: it exercises tools, prompts, resources and auth flows interactively, with a playground, multi-server chat and OAuth validation across four protocol versions, so you can check a server behaves identically across clients. README claims 16 client configurations and 170-plus models. Apache 2.0.

Snap Specs got another pitch in LA, this time wrapped in "Specs Intelligence," which Evan Spiegel describes as an AI native operating system that builds a model of your goals and routines from connected apps. HBO Max and Spotify apps, plus a Specs for Enterprise program with Amazon, Salesforce and NVIDIA. Still $2,200, cellular via Verizon at $10 a month for Verizon customers and $20 otherwise, shipping later this fall. Snap disclosed no battery, field-of-view or compute specs, which tells you something.

Thank you for reading today's edition.

Your feedback is valuable. Respond to this email and tell us how you think we could add more value to this newsletter.

Interested in reaching smart readers like you? To become an AI Breakfast sponsor, reply to this email or DM us on X!

Thinking of starting your own newsletter? AI Breakfast readers who sign up with Beehiiv receive a 14-day free trial and 20% off for 3 months.