Skip to content
Build a pack

AI Frontier Voices — Learning from the People Building It

What changed

This course is updated by hand, with a dated log. Newest first. Last updated 2026-08-31.

  1. Week of September 7, 2026

    Week of September 7, 2026

    • @polynoamial (OpenAI): The Astra that Brown teased in August is now a product. GPT-6 Astra shipped on September 3 to a limited set of organizations, then to Plus, Pro, Business and Enterprise users the next day, with the API (gpt-6-astra, $10/$50 per Mtok), Azure and Bedrock following. OpenAI's headline numbers: 99.9% on ARC-AGI-3, 97.6% on FrontierMath Tier 4, 72.6% on OSWorld 2.0 in roughly half the time of GPT-5.6 Sol, plus two new prime-gap results. Brown's framing is the architecture — a root agent that splits a problem across subagents running for hours or days — which is the "think for days, not fifteen minutes" thesis he has been building toward. Note the fine print: Astra hits OpenAI's Critical cyber threshold, so the launch version refuses proof-of-concept exploit work, and Enterprise access is off by default.
    • @karpathy, @bcherny & @trq212 (Anthropic): Claude Fable 5.1 and Mythos 5.1 launched September 1 — one model with two safeguard profiles, Fable generally available and Mythos gated to vetted cyber and life-science programs. List price is unchanged at $10/$50, but cache reads dropped 75% to $0.25 per Mtok, which Anthropic says makes typical work ~25% cheaper and cache-heavy agent runs up to 45% cheaper. Terminal-Bench Science more than doubled to 52.6%, and Claude Code users should see ~60% fewer cybersecurity false positives. Three breaking API changes for anyone upgrading: forced tool_choice now 400s, thinking blocks are one-directional across model versions, and editing history before a thinking block invalidates it.
    • @bcherny (Anthropic): Claude Code's v2.1.252–260 run made Fable 5.1 the default Fable model and shipped cross-session messaging: ListAgents discovers active sessions and SendMessage (or @name in a prompt) delivers text to one, over a local socket rather than Anthropic's servers, with per-session crossSessionInbound controls. Also new: /effort s, /code-review --comment for GitLab, and --permission-prompts none for headless runs. Calendar note: the +50% weekly-limits promo ends September 13 and is replaced by a permanent +25% on September 14 — a net ~17% drop from where you are today.
    • @OfficialLoganK (Google AI): Google shipped Gemini 3.8 Flash on September 2 — the third Flash in six weeks — with a Flash Cyber sibling for vulnerability detection and patching. Same $0.75/$3.75 pricing as 3.7 Flash (still doubling January 1), 1M context, and it beats Opus 5 on three published benchmarks. Read the caveat: it is 3.7 Flash tuned to burn more thinking tokens, and Google itself says to stay on 3.7 for efficiency-first workloads. Still no 3.5 Pro — seventh straight window, and the Pro line is now effectively skipped in every comparison table OpenAI and Anthropic published this week.
    • @milichab & @skcd42 (xAI): Musk posted on September 2 that Grok 4.7 "comes out in 10 days" — a September 12 target for the 2.1T model with the SpaceX engineering-data pass. No model card, pricing or context window yet; the docs still top out at grok-4.6. Around it: Grok Bot for Enterprise launched September 3 with isolated-VM agents, audit controls and a two-week trial, a read-only native X connector shipped August 29, and Grok Bot hit Android on September 2. Less flattering: Grok, Claude and ChatGPT all went down within 90 minutes of each other on September 3, with xAI blaming its Memphis compute center.
    • @leerob & @ericzakariasson (Cursor): Self-hosted machines shipped September 2: planning stays in Cursor's cloud while tool execution runs on your own laptop, VM or a team pool of workers, so code, build output and secrets never leave your network. Pools autoscale and hibernate, and cloud agents can now execute on AWS Lambda, Cloudflare, Modal, Vercel, E2B, Daytona, Coder and Namespace. Read alongside the OpenAI cutoff, Cursor is spending this month making itself harder to leave — infrastructure lock-in rather than model lock-in.
    • @mntruell (Cursor): OpenAI published its official statement on the Cursor cutoff, confirming the November 12 shutoff and that no future models (i.e. Astra) will reach Cursor. The stated reason is not commercial but trust: "we cannot be confident that SpaceX will use our technology within our terms of service." Anthropic, which rents compute from SpaceX, has kept its Cursor partnership — so on day one of GPT-6, the fastest-growing editor's model menu is Claude, Gemini, Grok and Composer.
    • @jxnlco (OpenAI): Codex 0.153.4 (September 4) makes Astra the bundled default model and adds the launch's most interesting Codex-specific feature: persistent notes across context windows, so long sessions keep searchable earlier context instead of lossy compaction (opt-in via config.toml, default "in the coming weeks"). Astra can also ask clarifying questions asynchronously in Codex while it keeps working on the parts that don't depend on your answer. Smaller: vim-mode search, scheduled time reminders, and an indexed web-search mode restricted to server-approved URLs.

  2. Week of August 31, 2026

    Week of August 31, 2026

    • @mntruell (Cursor): The week's biggest story. OpenAI notified SpaceX on August 28 that it will stop supplying models to Cursor on November 12 — and is withholding future models immediately — citing Elon Musk's history of breaking contracts. Truell responded on X that OpenAI models serve roughly 5% of Cursor traffic and that the teams are in talks. Two weeks after the $60B deal closed, the first real cost of being a SpaceX subsidiary arrived.
    • @jxnlco (OpenAI): A dense infrastructure week for Codex: August 24 deprecated the codex mcp-server command in favor of the app server, and August 28 added configurable MCP discovery grace periods, extension hooks to inspect or replace MCP tool results, and tighter sandbox and permission enforcement. The browser side widened too — the extension now runs in Edge, Brave, Opera and Vivaldi, with cloud browser sign-in for eligible plans.
    • @ammaar (Google AI): Introduced Gemini 3.5 Transcribe on August 26 — a speech-to-text model covering 85+ languages with smart correction and custom vocabulary — and, true to form, vibe-coded a Wispr Flow-style dictation app on top of it and open-sourced the demo alongside the launch.
    • @bcherny (Anthropic): Pricing housekeeping worth a calendar note: Claude Sonnet 5's promotional $2/$10 per Mtok ends August 31, with standard $3/$15 from September 1, and the +50% Claude Code weekly-limits promo is winding down in favor of a permanent 25% raise. Meanwhile the desktop app's built-in browser began rolling out in beta to Pro, Max and Team on macOS, Windows and Linux.

  3. Week of August 24, 2026

    Week of August 24, 2026

    • Anthropic platform: The agent stack went GA on August 20 — computer use, a new browser use tool, the Skills API and the Files API are all generally available on the Claude Platform. Browser use is genuinely new rather than renamed: it reads a page's accessibility tree and drives forms, tabs and uploads directly instead of screenshotting and guessing at coordinates. Computer use can now batch several actions in a single turn, and the Files API adds automatic expiration with 1 TB of org storage — if you have been carrying beta headers for any of these, you can delete them.
    • @bcherny & @trq212 (Anthropic): Claude Code's Week 34 digest (Aug 17–21) led with /design, a research preview that brings Claude Design's artboard workflow into the CLI and Claude Code Desktop on top of artifacts — Claude drafts editable artboards for your UI and implements the one you pick. Also shipped: a built-in Concise output style that skips preamble, device cards on your phone for any machine running claude remote-control, and ANTHROPIC_DEFAULT_MODEL to set the model new sessions start on.
    • Anthropic, education: Claude Academy launched this week — free courses, tutorials, learning paths, completion badges and an accompanying Claude Academy skill, spanning the consumer apps through the API, Console and MCP, with roughly 20 case studies and multi-language material. Worth noting for a course like this one: the labs are now building the curriculum layer themselves.
    • @leerob & @ericzakariasson (Cursor): Two shipments that read as one strategy. Origin (Aug 17, early beta on all paid plans) makes Cursor a code host — repos, pull requests, code browsing, GitHub sync. Then on Aug 19 cloud agents got Subscriptions: an agent subscribes to an event source (a PR, a Slack thread, a schedule) and wakes when something happens, auto-subscribing to PRs it opens and driving them through CI and bot comments. The same release added subagents on their own VMs, /goal, skills pinned as Custom Modes, and steering that waits for the next tool call instead of cutting the agent off. Cursor is going after the whole loop, not the editor.
    • @mntruell (Cursor) & @milichab (xAI): The merger stopped being paperwork. On August 21 xAI folded Grok Bot — its always-on agent teammates — into Cursor Pro+, Cursor Ultra and Cursor Teams plans, alongside SuperGrok Plus and Heavy. That is the first product where the SpaceXAI and Cursor sides share a SKU rather than compete, six days after the $60B deal closed. Modules 4 and 5 of this course are now describing one company.
    • @skcd42 (xAI): Grok Build went to every plan on web and mobile on August 19 — describe an app and Grok builds a working version live in the chat, publishes it to a grok.me link with a generated social card, plus remixing, custom domains, GitHub export, secrets and connectors. Grok 4.6's distribution also widened again: Amazon Bedrock on Aug 19 and Google's Enterprise Agent Platform Model Garden on Aug 21, both at $2/$6 per Mtok with a 500k context window. Grok 4.7 is still early September.
    • @polynoamial & @jxnlco (OpenAI): OpenAI cut GPT-5.6 Sol's API price by more than 20% on August 21 — $5 → $4 per Mtok input, $30 → $20 output — through at least November 21, covering pay-as-you-go, Codex credits and eligible ChatGPT Work plans. It is the second GPT-5.6 cut in under a month and it landed a week after Gemini 3.7 Flash undercut everyone. Consumer Plus/Pro/Business pricing is unchanged, so this is an inference-margin fight, not a subscription one.
    • @OfficialLoganK (Google AI): No Gemini 3.5 Pro, again. The week-of-August-21 window passed like the six before it, and the Gemini API changelog has published nothing since 3.7 Flash on August 13. Google's on-record line is still that it is testing with partners. Treat the 2M-token context window and Deep Think mode announced at I/O as unshipped until proven otherwise — nothing in the family offers more than 1M today.

  4. Week of August 17, 2026

    Week of August 17, 2026

    • @mntruell (Cursor): It closed. SpaceX officially completed the $60B all-stock acquisition of Anysphere on August 15, folding Cursor into its SpaceXAI unit with Truell's team moving into SpaceX's internal software engineering division. Cursor's own announcement leaned hard on compute rather than product, describing access to "the largest fleet of GPUs in the world" including Colossus — the same infrastructure SpaceX rents to Anthropic and Google. Module 4 of this course is now a story about a SpaceX subsidiary.
    • @OfficialLoganK (Google AI): Google shipped Gemini 3.7 Flash on August 14 — three weeks after 3.6 Flash — pitched as its "most intelligent workhorse model yet for coding and agents." It jumps to 43.6% on FrontierCode 1.1 Main (3.6 Flash scored 34.4%) and 1588 Elo on Code Arena for web dev, with a 1M context window. Read the pricing carefully: $0.75/$3.75 per Mtok is introductory through December 31, then doubles to $1.50/$7.50 on January 1. Kilpatrick attributes the gain to algorithmic improvements rather than compute — which is the only honest explanation available on a three-week cadence.
    • Still no 3.5 Pro: Worth stating plainly, since this is now the sixth straight window to pass. Google has shipped 3.6 Flash, 3.5 Flash-Lite, Flash Cyber and now 3.7 Flash since May while the Pro tier sits in limited preview. The Flash line has quietly become Google's actual frontier offering, and the interesting question is no longer when Pro ships but whether it ever does before Gemini 4.
    • @milichab & @skcd42 (xAI): Grok 4.6's distribution widened fast — it landed in GitHub Copilot on August 14 across eight surfaces (VS Code, Visual Studio, Copilot CLI, cloud agent, the Copilot app, JetBrains, Xcode, Eclipse) on Pro through Enterprise SKUs, alongside OpenRouter, Vercel and Cloudflare. Note the deadline: the doubled-usage promo in Cursor and Grok Build closes August 19. Meanwhile Grok 4.7 slipped to early-to-mid September, and the API model list still tops out at grok-4.6.
    • Anthropic platform: The Claude in Chrome side panel became a full Cowork session on August 12 — history, skills and connectors all work inside the browser, and a task started in a tab can be finished on desktop, web or mobile. Max and Team have it now; Pro is rolling out over the coming weeks, and Enterprise is off by default with admin domain allowlisting. This is the first time the browser stopped being a separate surface and became just another place the same session lives.
    • Anthropic, provenance: Coverage caught up this week to the invisible watermarking of all Claude text output, live since August 2 and applied worldwide rather than only in the EU, in response to Article 50 of the EU AI Act. It works by statistically biasing word choice against a key Anthropic holds — undetectable per sentence, detectable across enough text — and travels through copy-paste and some editing. Files get C2PA provenance metadata instead, which is strippable by format conversion, re-saving, or a screenshot. If you ship Claude-generated text or assets, this now applies to Claude Code and Cowork output too.

  5. Week of August 14, 2026

    Week of August 14, 2026

    • @bcherny (Anthropic): Auto mode flipped to default in Claude Code today for new sessions on Pro, Max, and Team — the change trailed last week from the 1,053-tester classifier study. Enterprise, the Claude API, and the cloud platforms (Bedrock, Vertex, Foundry) stay opt-in, with Anthropic saying it plans to make it default there within a month. If you script Claude Code, assume permission behavior changed underneath you.
    • Anthropic platform: Two enterprise pieces landed alongside it. Self-hosted environments went to public beta (Aug 6) — claude self-hosted-runner turns your own machines into the compute layer, so agents sit next to private registries and databases without exposing them. Read the fine print: it is self-hosted execution, not self-hosted inference — prompts still go to Anthropic. Separately, the Compliance API now covers Cowork and Claude Code sessions with consolidated transcripts.
    • @karpathy (Anthropic): Weighed in after Claude Opus generated a procedural 3D Lord of the Rings world from a single prompt, arguing models have moved past the prompt-a-snippet era into sustained creative construction. It is the same thesis as his loop-engineering post, just with a demo attached rather than a rulebook.
    • @leerob & @ericzakariasson (Cursor): Cursor shipped Builds on August 13 — pre-warmed copies of your dev environment kept ready in the background, so a Cloud Agent boots into a working setup instead of provisioning one per session. Cursor claims 10× faster environment boot and 3× faster time-to-first-token, and it is included with Cloud Agents at no extra cost. Environment setup was the single biggest tax on cloud agents; this is the right thing to have fixed.
    • @mntruell (Cursor): SpaceX finalized regulatory procedures on August 12 to close the $60B all-stock acquisition of Anysphere, keeping the Q3 timeline. The joint SpaceX/xAI/Cursor model intended to ship inside both Cursor and Grok is the thing to watch — it would make Cursor and Grok Build siblings rather than competitors, which reshapes module 4 and module 5 of this course at the same time.
    • @milichab & @skcd42 (xAI): With 4.6 out, attention moved to Grok 4.7 — now roughly 3–4 weeks out as of August 12. Musk says initial training is done and xAI is running a supplemental pass on "a massive amount" of SpaceX engineering data, and frames 4.7 as better than 4.6 in every way except slightly slower to serve, with better token efficiency. Training on proprietary aerospace engineering data is a genuinely new kind of moat, not just more scale.
    • @OfficialLoganK (Google AI): The August 12 rumor for Gemini 3.5 Pro passed with another delay, this time attributed to deployment issues rather than quality bars. That is now five missed windows since May. At some point the interesting question stops being "when does 3.5 Pro ship" and becomes "does Google skip it for Gemini 4."
    • @gabriel1 (OpenAI): Calendar item for anyone still on Sora: the app shut down back in April, and the Sora 2 API is scheduled for removal on September 24 in favor of Spud. Six weeks left to migrate any video pipeline built on it.

  6. Week of August 9, 2026

    Week of August 9, 2026

    • @milichab & @skcd42 (xAI): Grok 4.6 shipped on August 7, on schedule and to the walked-back spec: 1.5T parameters on the same V9 foundation as 4.5, with the gains coming from post-training rather than scale. It went out through xAI's own surfaces first — SuperGrok tiers, the API console, and Grok Build. The 2.1T Grok 4.7 is still a few weeks out.
    • @bcherny (Anthropic): Anthropic announced that auto mode becomes the default in Claude Code on August 14 for Pro, Max, and Team. The number that carried the argument: across 1,053 paid testers, auto mode's classifier caught 89% of harmful actions versus 13.6% for humans clicking through manual permission prompts. Anthropic also stopped charging for the classifier's token overhead. Enterprise, API, and cloud-platform users stay opt-in for now.
    • @polynoamial (OpenAI): The Astra verdict firmed up over the week. Thomas Bloom — the mathematician who took apart OpenAI's 2025 math claim — called this result "big news" and rated it above the May 2026 Erdős work, which is meaningful validation from the most credible skeptic available. OpenAI has also started framing Astra less as a math model and more as a long-horizon system that coordinates multiple agents over extended runs.
    • @OfficialLoganK (Google AI): The August 6 window for Gemini 3.5 Pro came and went with no launch — the model is still in limited preview on Vertex AI. That's the fourth slip since the May 19 I/O announcement, and the new unconfirmed rumor points at August 12. Meanwhile Gemini 4 pre-training continues, which increasingly looks like the real explanation.
    • @leerob & @ericzakariasson (Cursor): Cursor restored the Cost column and historical CSV dollar values on August 2 after replacing them with raw token counts drew user pushback — a small thing, but a clean example of the pricing-transparency pressure every agent product is under. Cursor for iOS also moved to public beta on all paid plans, and a lower-priced Cursor Start tier launched for developers in India.
    • @jxnlco (OpenAI): Worth a calendar note for anyone deep in Codex: GPT-5.4 and 5.4-mini stop being available in Codex for ChatGPT-authenticated sessions on August 31, migrating to the GPT-5.6 Terra and Luna variants. API-key sessions keep access. If your workflows pin a model, they will break silently.

  7. Week of August 2, 2026

    Week of August 2, 2026

    • @polynoamial (OpenAI): The week's biggest story. Brown announced on August 1 that an internal version of Astra — OpenAI's next major model family — produced new results on ten problems open for at least a decade, across sphere packing, von Neumann algebras, circuit complexity, quantum complexity, lattice cryptography and Ramsey theory. Each came with a machine-checkable Lean certificate, and the whole run cost roughly $2,000 in API compute. Brown's own caveat matters as much as the claim: "sadly no Millennium Prize problems (yet)… we didn't spend a lot on each problem."
    • Counterpoint worth reading: Gary Marcus published "amazing — but vastly oversold" on August 2, arguing the methodology isn't public and that math is a special case: it has symbolic verification and cheap correct synthetic data, so success there doesn't transfer to open-ended reasoning. He also flagged that at least one Astra proof may be wrong. Hold both — the result is real and the generalization claim is unproven.
    • @milichab & @skcd42 (xAI): Musk dated Grok 4.6 to around August 7 on July 28, and the specs walked back last week's numbers: 1.5T parameters on the same V9 foundation as 4.5, not a 2T new base. The improvement is post-training — better SFT and RL — with 4.7 (2.1T) a few weeks out. xAI is now shipping gains from training technique rather than scale, which is a notable shift in its own right.
    • @trq212 (Anthropic): His context-engineering post for Claude 5 models kept spreading all week. Anthropic cut Claude Code's system prompt from ~2,686 words to ~514 with no measurable loss on coding evals — but the instructions were relocated into tool descriptions and skills, not deleted. The six shifts he names (rules → judgment, examples → interface design, upfront context → progressive disclosure) are the most directly actionable thing on this list for anyone writing CLAUDE.md files.
    • @karpathy (Anthropic): Released a document on agent architecture that trended on X — nine rules for pushing execution and bookkeeping onto the model and keeping state on disk, so the human sets up a gen-verify loop instead of prompting turn by turn. It's the formal version of the "agentic engineering, not vibe coding" line he's been making since Sequoia Ascent.
    • @OfficialLoganK (Google AI): Gemini 3.6 Flash and 3.5 Flash-Lite went generally available in the API, and Kilpatrick has now hinted at an August window for 3.5 Pro after three slips. Four months of Pro delay while Gemini 4 pre-training is already underway is starting to look less like a schedule problem and more like a decision about which model is worth shipping.

  8. Week of July 26, 2026

    Week of July 26, 2026

    • @OfficialLoganK (Google AI): Google shipped three Gemini models on July 21 — and none of them was 3.5 Pro: Gemini 3.6 Flash ($1.50/$7.50 per Mtok, 1M context, ~17% fewer tokens than 3.5 Flash), 3.5 Flash-Lite ($0.30/$2.50), and a gated 3.5 Flash Cyber tuned for finding and fixing vulnerabilities. Kilpatrick said 3.5 Pro is still testing with partners. The footnote was the real news: Google has begun pre-training Gemini 4.
    • @karpathy (Anthropic): A bio tweak on X plus Anthropic's absence from Jensen Huang's open-weights letter spun up a "Karpathy quit Anthropic" rumor, which he flatly denied on July 26. He remains on pre-training under Nick Joseph. Worth noting how fast a profile edit became a narrative — a reminder to read primary sources on this list, not the repost layer.
    • @milichab & @skcd42 (xAI): Musk announced Grok 4.6 — a 2T-parameter model — finished its initial training run the week of July 20, with a 4–6T Grok 4.7 already on the roadmap weeks behind it. That's a back-to-back cadence no other frontier lab is matching. xAI also pushed Grok into Google Workspace as a free Sheets/Slides/Docs add-on with cited-cell answers.
    • @ammaar (Google AI): Reshi compiled Command & Conquer: Generals — Zero Hour to run natively on iPhone, iPad and Apple Silicon — no emulator — using Claude Code with Fable 5, then open-sourced it. First build in ~40 minutes, roughly two days total. The AI Studio product lead shipping his best demo on a competitor's tool is its own signal about where agentic coding actually stands.
    • @jxnlco (OpenAI): Published his full system for running an entire workday inside Codex (July 26) — a "chief of staff" thread that handles Slack and email, plus long-running tasks defined against verifiable goals. The through-line of his Codex-maxxing series: the win isn't code generation, it's giving an agent work that can be checked when it finishes.
    • Ecosystem context: Moonshot's Kimi K3 open weights land July 27 — 2.8T parameters, sparse MoE, 1M context, Modified MIT. It benchmarks above Grok 4.5 and is the immediate target Grok 4.6 was built to beat. The catch: ~1.4TB of fast memory even at four-bit, so "open" here means open to whoever owns the hardware.

  9. Week of July 19, 2026

    Week of July 19, 2026

    • @bcherny (Anthropic): Published "Steps of AI Adoption" (July 16), naming five maturity levels — Gated (0) → Assisted (~1) → Parallel (~10) → Supervised autonomy (~100) → AI-native (1,000+ agents). His argument: the 10× gap between teams isn't "more tokens," it's bottlenecks and guardrails at each step, so companies should measure human hours saved, not token burn.
    • @skcd42 (xAI): xAI open-sourced Grok Build under Apache 2.0 (July 15) and made Grok 4.5 its default model. The terminal agent runs up to 8 parallel subagents in isolated git worktrees; xAI also disabled a data-upload feature after community security feedback. Grok Build going open source is a direct shot at Claude Code and Codex.
    • @leerob & @ericzakariasson (Cursor): Cursor doubled included model usage on all plans (July 16), widening access to Grok 4.5 and Composer 2.5. Cursor in Slack (July 17) now shares a plan before it executes and works across multi-repo environments and threads — agents you can redirect early rather than after the fact.
    • @karpathy (Anthropic): His nano series (nanoGPT, nanochat, micrograd) crossed a combined ~120,000 GitHub stars, cementing him as ML's most-followed educator. Anthropic's hiring run continued alongside his pre-training work — Monzo co-founder Tom Blomfield joined the compute team.
    • @OfficialLoganK (Google AI): Gemini 3.5 Pro slipped again — the widely-reported July 17 launch passed with no release after the rebuilt model reportedly missed Google's internal quality bars (its third delay). AI Studio still surfaces gemini-3.5-flash and a 3.1-pro-preview; the new Pro (2M-token context, "Deep Think" reasoning) remains in limited enterprise preview.