9 min read · 14 sources

OpenAI just published the kind of internal memo competitors usually only see in court filings: a candid look at its progress on recursive self-improvement, the loop where models help build and evaluate the next, more capable models. The company frames AGI as something that must be democratically governed, and says public debate is the only way that happens. The provocative part is the call to action: OpenAI wants other labs to be required to publicly track their own RSI work, not just release snapshots of finished models.

It is hard to read that as anything other than a request for regulatory disclosure aimed at peers. The argument is that capabilities, risks and safeguards only get debated honestly if the underlying numbers are visible. For engineers, the practical upshot is that the frontier-lab race may start producing benchmarks and progress reports on the metaresearch loop itself, not just on model outputs.

A Tesla Cybercab in Austin priced an 8-mile ride at $14.87 against $25 on UberX, but the wait was 30 minutes versus 2.

Tim Cook's pay package doubles down on him staying

Source: bloomberg.com ↗

Apple has handed Tim Cook a base salary of $2 million, roughly twice the typical SVP base in the C-suite, plus cash bonuses and vesting stock units. The size of the package is itself the story: Bloomberg’s read is that Cook is being paid as the active CEO he is, especially for the parts of the job no one else wants, including US-China trade handling and the ongoing memory shortage.

The memory squeeze in particular makes Cook’s relationships with TSMC, Samsung and SK Hynix load-bearing. A package this rich is a board signalling it does not want a transition during a constrained component cycle.

Travis Kalanick's Atoms is now a robotaxi play

Source: techcrunch.com ↗

Atoms, the Travis Kalanick startup that raised a $1.7 billion round led by Andreessen Horowitz earlier this summer, is gearing up for a hiring and acquisitions spree aimed at autonomous vehicles. Uber has put $100 million in and is already talking about using Atoms’ robotaxi tech inside its own network. The trajectory rhymes with Atoms’ earlier pickup of Pronto, Anthony Levandowski’s autonomous mining startup, which itself is a tell: Kalanick is buying into autonomy, not building it from scratch.

For the rest of the robotaxi stack, this matters. Uber brokers partnerships with Waymo and Cruise; Atoms is now a well-funded alternative negotiating from inside the same relationship. Whoever wins that auction will determine whose perception, mapping and dispatch software ends up under a million ride-hail trips.

Why AlphaProof is less impressive than it looked

Source: contraptions.venkateshrao.com ↗

A long essay from Venkatesh Rao reframes DeepMind’s AlphaProof and its 2024 IMO success (3 of 5 non-geometry problems solved) as an artefact of mathematics being turned into a game over the past century, not as a general reasoning breakthrough. The technical bones: Lean encodes a proposition as a type and a proof as a term of that type, with a small kernel that mechanically verifies proofs; Mathlib supplies a vast library of definitions, lemmas, and tactics; LeanDojo exposes proof states as inputs so a language model can generate tactics and treat premise selection as retrieval.

Put those together and you have states, moves, an accumulating body of knowledge, a terminal condition, and a verifier, exactly the shape a learning system needs. The argument is that frontier results in formal math are evidence of very good domain engineering, not evidence that the same model can reason in domains where the rules are fuzzier. The framing is uncomfortable: a lot of what looks like “general AI” right now is really “we made this domain look like a game.”

Handing production secrets to coding agents

Source: jacob.gold ↗

Jacob Gold writes that he now gives coding agents direct access to live secrets: his Tailscale auth key, SSH private keys, Anthropic and Codex OAuth tokens, and a Gmail app password. He accepts Simon Willison’s “lethal trifecta” risk (private data + untrusted content + outbound comms) on the bet that frontier models (he names Astra and Fable for hard work, Opus and Sol for routine) are much harder to prompt-inject than older ones.

He still wraps each agent in a per-task Docker container, but explicitly notes that containers are not his security boundary. The post reads as both a recipe and a warning: the productivity delta from giving agents root is large enough that he has decided the blast radius is worth it.

Terraform, but the errors land at compile time

Source: ngrislain.github.io ↗

An engineer has built infra, a Terraform-like IaC tool written in Lean 4. Roughly 15,000 lines of Lean across 107 commits over two weekends, covering 3 clouds and 14 resource kinds, and he is upfront that it is not production-ready. The pitch is the type system: a region literal like warsaw compiles only when every cloud in scope recognises it; the moment AWS enters the module, that string stops compiling because AWS has no Warsaw region.

The post tabulates the rest of the wins, including missing references, unmet dependencies, plans that depend on unreturned values, nonexistent instance sizes, and unsupported services, all caught before any apply. Bucket name collisions, quota, capacity and eventual consistency remain runtime problems. The dependent-types approach is the same idea Pulumi and CDK try to give you, except it is enforced by a real kernel instead of a linter.

Brockman on Astra, alignment and OpenAI's place in the stack

Source: stratechery.com ↗

Stratechery has a long interview with Greg Brockman, recorded before the Astra launch. Brockman walks through his path from the International Chemistry Olympiad to Stripe’s CTO seat to co-founding OpenAI in 2015, the ChatGPT launch, the 2023 leadership drama, and where OpenAI sits between Microsoft and Nvidia on the AI value chain.

The interview also revisits the “Hugging Face incident,” debating whether OpenAI took security seriously in the run-up to it. For engineers tracking the launch, the interesting bits are the constraints Brockman names around scaling and the alignment commitments the company is willing to make on the record.

The agent loop is closing inside real companies

Source: henryaj.substack.com ↗

A first-person account from Henry Aj builds on Karpathy’s 2025 framing of three generations of software (handwritten 1.0, narrow ML 2.0, LLM-written 3.0) to argue org processes are being automated by agents the same way codebases were. Concrete examples from his own team: Claude picks up Sentry errors from Slack, pulls logs, runs diagnostics and opens a change request, with the only human step being the merge click; product requests come in via Slack, get specced with the user, coded and ticketed; staging deploys are QA’d by Claude driving a browser, including visual regression checks.

The thesis is uncomfortable: companies 1.0 are being eaten by agents that can be parallelised indefinitely and never sleep. The interesting constraint Aj flags is context: humans are still better at absorbing, parsing and sharing context with each other. The day that gets automated is the day the loop closes for real.

A fruit-fly connectome recognises human emotion

Source: oruk.ai ↗

Researchers at oruk.ai wired up 499 neurons from the MaleCNS fruit-fly connectome (15,865 connections, 867,344 synaptic contacts) as a fixed recurrent reservoir and trained a tiny linear readout to recognise 15 emotion labels and 16 speaking styles from human speech. Audio is sliced into 32 frequency-band energies every 10 ms, projected into the circuit, and the fly’s wiring is left frozen while only the readout learns.

It sits inside reservoir computing, citing Suárez et al.’s conn2res and earlier work by Costi et al. using fly connectomes for time-series prediction. The point for ML engineers is a clean number: how much downstream signal a fixed biological recurrence contributes with zero learned internal weights.

Cybercab vs UberX, same Austin morning

Source: farzad.fm ↗

Farzad’s day-two tape from the Austin geofence: an ~8-mile Domain-to-UT ride cost $14.87 on Cybercab with a ~30-minute wait, $19.68 on Model Y with a ~7-minute wait, $25 on UberX in ~2 minutes, $32 Comfort, $56 Premiere. Public Cybercab rides launched Sept 4 at 5 p.m. CT and were moved up to 2 p.m. CT; he estimates ~40 Cybercabs and 10 - 12 Model Ys in the geofence, with Texas DMV listings showing 45 Cybercabs on Tesla’s statewide roster.

The thesis is that once fleet density rises, a no-steering-wheel, no-pedal stack lets Tesla undercut human-driven margins. Right now the wait is the tax.

iPhone Ultra production is running at "a few hundred" a day

Source: macrumors.com ↗

Nikkei Asia reports Apple’s foldable iPhone Ultra was producing only “a few hundred units a day” in late August, held back by strict QC on surface flatness and hinge performance despite a raised annual order of ~10 million units (up from 7 - 8 million) following Foxconn trial production. An extra verification step in August pushed commercial production back “weeks,” raising the probability of a staggered launch that starts in the US. Apple is still expected to unveil it alongside the iPhone 18 Pro and 18 Pro Max at its September 9 event.

The hinge and display durability bottlenecks are the real story; they are why foldables have been hard to scale for everyone, not just Apple.

Inside the Apple Neural Engine compiler

Source: smolnero.com ↗

A reader’s digest of Spencer Bryngelson’s reverse engineering of the Apple Neural Engine compiler, set against the better-documented LLVM, KGEN, MLIR and Mojo stacks. Apple exposes the ANE only through Core ML; the compiler, program format, driver and firmware are undocumented, so the paper is built on direct measurements of Apple’s private software.

The author walks a simple convolution → bias → ReLU graph through the lowering pipeline and shows that, contrary to the naive expectation, these three ops do not simply map to three lower-level ops on the accelerator. It is one of the few public sources sketching how Core ML graphs end up on ANE silicon.

Leaving the Mac behind for Omarchy

Source: hadijaveed.me ↗

After 12 years on macOS, one engineer moved off the Mac because an M4 MacBook with 48 GB of RAM was struggling under multiple Claude Code, Codex and Docker sessions and needed daily restarts. He moved his dev environment and agents to a server (Codex remote, Hermes, Ghostty, Herdr) and replaced the laptop with a Lenovo Yoga Slim 7i Ultra Aura Edition running Omarchy, DHH’s opinionated Arch + Hyprland distro, initially via a Beelink SER8 mini desktop.

He highlights keyboard-driven tiling (Rectangle, AeroSpace and Raycast on Mac; native tiling in Omarchy) and the Quattro plugin ecosystem, including plugins that surface Claude and Codex usage limits and Tailscale peers. The pattern is concrete: thin Linux client + remote agent host, with the heavy lifting no longer on the laptop.

GPT-6 Astra and the "fourth exponential"

Source: taekim.substack.com ↗

Tae Kim argues GPT-6 Astra is a milestone, citing reactions like Helion Capital’s Houman Dehesh saying “Astra devastates Fable” on financial data, and Greg Brockman’s claim that computer use has crossed a qualitative threshold from slow and clumsy to accurate and reliable. Kim tested it himself by asking Astra to build a detailed Space Shuttle model in Blender (a program he had never used) and watched it research the topic, write scripts and drive Blender on its own, producing a rotatable 3D model in ~10 minutes.

His read is that computer-use AI is now “real” and will soon outperform most humans at software knowledge work. The framing of a “fourth exponential” is the bet: demand for AI silicon has surged on chatbot, reasoning and agentic-coding waves, and computer tool use is the next one.