Good morning. It's Monday, September 14th.

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

Amodei says pace the frontier. Every other frontier CEO said yes within a day.

Amodei published "We Must Pace the Frontier" on Saturday. The argument starts with a claim about the last three months: "since roughly this summer, AI has been advancing drastically faster, driven primarily by AI's growing ability to build the next generation of AI." His main fear is agent swarms. He writes that "in 6 to 12 months such a swarm could be capable of taking over the entire internet with a persistent botnet (potentially causing hundreds of billions of dollars in damage)."

Dario’s plan has three steps, and only the first is something Anthropic can do alone. Embedded third-party evaluators get, in his words, "Desks in our offices, access badges, and company laptops," plus permissions "mostly comparable to what internal risk assessment teams have," and the right to publish findings "without editorial control by Anthropic." Step two is common standards and capability limits across democratic-world labs, which he says needs Washington to "issue a narrow waiver for certain kinds of safety conversations" because of antitrust. Step three is coordination with authoritarian governments, which he concedes is the hard one. On China he asks for no new limits on US labs at all: keep chip export controls, crack down on distillation, harden weight security.

Sam Altman posted that he agrees "we need to pace the frontier," called it "a primary topic of discussions we've had at OpenAI in recent weeks," and said OpenAI will match the independent-evaluator commitment. Musk replied "Dario is right." Nadella backed pacing on Sunday and said Microsoft would open its first-party MAI model behavior rules to public consultation. That is four of the largest labs on the same side of an argument inside two days. Personally, I think it’s a bullshit face-saving maneuver and none of these companies intend on slowing development.

Read more.

The market read "pace the frontier" as a capex cut

Memory took it worst. SK Hynix closed down 6.4% and Kioxia fell 6.4%, Samsung Electronics lost 4.1%, and the KOSPI dropped roughly 3.2% in early Monday trade. Europe followed at the open: as of 07:10 GMT, ASML was off 4.4%, BE Semiconductor 4.8%, Infineon and ASM International both more than 5%, STMicroelectronics 3.5%. SoftBank fell as much as 13%, its worst session since July. US names barely moved by comparison, with Nvidia and Micron roughly flat and Arm up.

Bernstein's Stacy Rasgon argues the reaction overshoots the text. Semis were already down about 19% from June peaks, and as his team put it, Amodei "is not calling for a halt to training, but rather a shift from 'extremely fast' to 'only somewhat fast.'" The essay asks for no compute cap on US labs. It asks for evaluators and export controls. Bernstein still likes Nvidia, Broadcom and the equipment makers.

The timing is the part worth sitting with. Amodei publishes a slow-down manifesto Saturday. The Financial Times reports Anthropic has told shareholders it expects a second straight quarter of positive adjusted operating income on gross margins above 80%, against $11.5B in Q2 revenue. Business Insider reports Anthropic has picked Nasdaq for an October listing, with Reuters separately reporting talks for Nvidia to anchor with up to $10B. On Monday, Bloomberg reported SoftBank closed an upsized $11.87B two-year loan from roughly 20 banks to fund its OpenAI investment. Altman, meanwhile, told Fortune on Friday that safety makes right now "an ill-advised moment to go public," and confirmed: "I would say not 2026, yeah." Every one of the Anthropic financial figures is single-outlet reporting from unnamed sources with no company confirmation, so hold them loosely. The direction of the calendar is not in dispute.

Read more.

A benchmark built on private production code puts the best model at 38.8%

While everyone argued about how fast to go, a small shop called Specific published Real-SWE, which swaps public GitHub issues for licensed private codebases: a social events app with 200,000-plus users, a fintech platform processing 100,000-plus bank statements, an enterprise AI sales product. Tasks touch roughly 11 files each, against about 6 on comparable public benchmarks.

Nobody passes. Fable 5.1 resolves 38.8%, GPT-6 Astra 33.8%, Gemini 3.8 Flash 31.2%, GLM 5.3 28.8%, Grok 4.6 and Muse Spark 1.3 both 23.8%, Kimi K3 18.8%, GPT-5.6 Sol 16.2%. That is a long way from the 90-plus Terminal-Bench numbers these same models post.

Caveat it properly: 10 tasks, 8 runs each, 640 scored rollouts total, and the harness and dataset are not open, only sample tasks on request. That is a small sample from an interested party, and it is not a replacement for a public benchmark. It is still the cleanest available argument that the gap between benchmark scores and production engineering is enormous, and it landed on the weekend the industry decided capability growth is the emergency.

Read more.

Rune is a GPU-rendered, keyboard-driven IDE written in Go that bundles editor, terminals, language intelligence and debugging, and deliberately keeps the AI agent outside the core as a separate extension. Now GPLv3 and free to build from source.

buildprof traces every process a Linux build spawns via ptrace and renders the whole thing as a Perfetto timeline. The author pointed it at Bun and found Full LTO linking, not the language, eating 16-plus minutes. Apache 2.0.

Cortex takes one config file over OpenAPI, AsyncAPI, GraphQL SDL, Protobuf and OpenRPC and emits typed SDKs in 11 languages, interactive docs, and an MCP server so agents can call the API. MIT.

Litelm strips LiteLLM's actual call path (routing, message translation, streaming, tool use, embeddings) down to about 2,300 lines with two dependencies, against LiteLLM's 100k-plus. MIT, and explicitly alpha.

ClinePass is a $9.99 per month subscription bundling 16 open-weight models across six providers, including GLM 5.3, Kimi K3, DeepSeek V4 Pro and MiniMax M3, at what Cline claims is 2x to 5x standard API rate limits. Core runtime stays Apache 2.0.

Homebrew 7.0.0 adds brew vulns vulnerability scanning against an advisory database, overlapped download and prep, stronger sandboxing and a native macOS BrewUI app. Intel Macs drop to Tier 3, with no new bottles after September 2027.

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.