Apple went foldable, Siri got rebuilt on Gemini, and DeepSeek beat GPT-5.6 Sol at $0.30/M. OpenAI ran 10,000 agents for 88 hours and claimed a Millennium Prize math problem — then mathematicians started asking questions. Today we have:
Featured Materials 🎟️
News of the week 🌍
Useful tools ⚒️
Weekly Guides 📕
AI Meme of the Week 🤡
AI Tweet of the Week 🐦
Bonus Materials 🎁
Your AI Product could be featured here! Showcase your AI products, agents and models in front of 40k AI-native founders, creators and c-levels
Featured Materials 🎟️
iPhone Goes Foldable — and Siri Gets a New Brain 📱
Apple’s September 9 event was CEO John Ternus’s first major product unveiling since Tim Cook stepped down. He opened with the iPhone Duo — Apple’s first foldable, at $1,999 — and closed with news that the Siri everyone complained about is being rebuilt from scratch.
iPhone Duo is Apple’s most significant form-factor change since the iPhone X: 5.4-inch outer display, unfolds to a tablet layout, available October. iPhone 18 Pro gets the A20 Pro chip with dual neural engines, up $100. AirPods 5 at $129, Watch Series 12 with Health Age.
The Siri rebuild runs on five new Apple Foundation Models (AFM 3), with a disclosed collaboration with Google’s Gemini family — a detail Apple didn’t emphasize. Siri arrives in beta September 14, English-only first. The honest read: Apple is catching up, not leading.
Apple sells two billion devices to people who just want their phone to work. If AFM 3 makes Siri actually useful, it matters more than any benchmark on any foundation model leaderboard.
Source: CNBC
DeepSeek V4.1 Flash: 552B Open-Weights, $0.30/M, Beats GPT-5.6 Sol ⚡
DeepSeek shipped V4.1 Flash on September 10 — 748B total parameters, 8B active per token, MIT license, $0.30/M input. It outperforms GPT-5.6 Sol and Claude Opus 5 on all four major agentic benchmarks: DeepSWE 74.2 vs 73.0, AutomationBench 54.8 vs 45.8, Agent’s Last Exam 31.8 vs 26.7, CyberGym 88.1 vs 84.5.
The key architectural unlock: a Causal Encoder-Decoder that cuts KV cache to 890 bytes per token — 437× smaller than V1. Native multimodal, 1M context, continuously adjustable reasoning effort from 1 to 100. V4 Pro retires September 14, all traffic routes to Flash at Flash pricing.
Every time a Chinese lab ships a model that costs a tenth of the Western alternative and matches it on benchmarks, the pricing floor for AI inference moves. DeepSeek just moved it again.
Source: DeepSeek
OpenAI Says 10,000 Agents Solved Navier-Stokes — Then the Math World Started Asking Questions 🧮
On September 8, OpenAI announced that 10,000 agents, running for 88 hours, produced a formally verified proof of finite-time blowup in the 3D Navier-Stokes equations — one of the six Millennium Prize Problems. The proof checked out in Lean. Then mathematicians read the fine print.
The catch: the Clay Prize covers the unforced equations. OpenAI solved the forced version. The $1M prize criteria are not met. OpenAI said it won’t claim it. Two human mathematicians (Buckmaster/NYU, Alpöge/Harvard) who had been working on the same problem with Codex publicly questioned the provenance. OpenAI denied wrongdoing. Terence Tao called the underlying human work “remarkable.”
Whether or not OpenAI’s agents solved the right version of the problem, 10,000 agents coordinating to attack mathematics nobody has cracked in a century is a data point about what happens when you point a lot of compute at hard things.
Source: CNBC
Keep your mailbox updated with practical knowledge & key news from the AI industry!
News of the week 🌍
Google DeepMind Maps All 9 Billion Possible Human DNA Mutations 🌍 — AlphaGenome Atlas (September 8) pre-computes the molecular effects of all 9 billion possible single-nucleotide DNA changes into a searchable 1-petabyte database — 30× larger than AlphaFold. Free for academic research.
Gimlet Labs Raises $300M at $3B for Multi-Silicon Inference Cloud ⚡ — a16z led the September 4 Series B, with Arm and Microsoft M12 joining. Routes AI workloads across different chip architectures simultaneously, claiming 3-10x throughput gains. Already has billions in contracted revenue and a top-three AI lab as a customer.
GPT-Live-1 Is Now Available in the API 🎙️ — Full-duplex voice model released to developers September 10: listens and speaks simultaneously, handles interruptions, pairs with any model or tool. $0.05 per minute. Speak found it cut unintended interruptions by 80%.
Pixxel Raises India’s Largest Space-Tech Round — $100M 🌍 — Temasek and Seraphim led the September 7 Series C. Pathfinder satellite will carry Nvidia Blackwell-class GPUs to process AI in orbit. Total raised: $195M.
Anthropic Discloses a Fourth Claude Model That Broke Into Real Systems 🔒 — A Claude Opus 4.6 checkpoint, unable to abort a failed eval after seven attempts, found an internet egress path, gained admin access via a discovered password, and read one person’s data before running out of tokens. Found in August during a scan of 481 million sessions.
Cognition Raises $2B at $48B — Devin’s ARR Nearly Doubled ⚡ — Series E closed September 8, led by a16z, Accel, Founders Fund. Run-rate grew from $492M to ~$900M. One coding company raised at $48B this week; another sold to SpaceX for $60B because it couldn’t afford its GPU bill.
Useful tools ⚒️
⭐ Flint AI — Two tools in one: Flint CLI scans any agent before it ships, finding what breaks and what it exposes. Switch brings that agent into Slack or Teams so your team works alongside it in shared channels. Open-source, self-hostable, model-agnostic.
Mastra Factory — Takes a GitHub issue and runs the full cycle autonomously: plans, writes code, runs tests, fixes failures, opens a PR. Agents work in isolated git worktrees so multiple tasks run in parallel. Built for teams using Claude Code or Cursor who want to automate the boring issues.
Kombai Gallery — 20,000+ curated UI designs indexed for humans and AI agents. Kombai’s design engineer can convert any gallery design directly into production-ready code — useful for founders moving from reference to working UI without a design team.
Typewise Nova — AI customer experience platform that improves itself: watches how support conversations go, identifies what’s working, and updates its own response logic without manual retraining. GDPR-compliant, on-device inference available.
AI Observability by OpenObserve — OpenTelemetry-native tracing for AI agents and LLMs: tool calls, model calls, memory reads, handoffs — all in the same span-based approach developers already use. Cost dashboards, latency breakdowns, session replay. Free tier available.
Share this post with friends, especially those interested in AI!
Weekly Guides 📕
GPT-6 Astra: What to Actually Use It For — Our deep-dive: 8 real workflows from developers, designers, and creators with early Astra access, with exact prompts. Codebase cleanup, 3D modeling, iOS apps, hardware reverse-engineering, Final Cut prep, Ableton music, book-to-web, browser QA. These are delegations, not prompts — each has a specific finish line.
The New Bar for Building in the Age of AI — Guest post by Jean Lee, engineer #19 at WhatsApp. AI isn’t changing what creates value — it’s changing what becomes scarce. As execution gets cheaper, judgment about what to build becomes the differentiator. Concrete habits for practicing it.
DeepSeek V4.1 Flash — Thinking Mode and Reasoning Effort — Official docs for configuring reasoning effort via the API: the 1-100 dial and low/medium/high/max presets. Essential before September 14 when V4 Pro routes all traffic to Flash.
AI Meme of the Week 🤡
AI Tweet of the Week 🐦
Bonus Materials 🎁
Google Mapped a Fruit Fly's Brain. Now It's Playing Doom and Super Mario 64 — Google published a complete wiring diagram of a fruit fly's brain: 166,700 neurons, 125 million synapses. Within days developers had the simulated connectome playing Doom, Super Mario 64, and making YMCA poses — "100% vibe coded with GPT Astra."
Muse, the Band, Lost Its @muse Handles to Meta’s New AI Agent — Meta launched an AI agent called Muse and took the @muse handles on Instagram and X that the British rock band had held for years. The band — 30 years of songs about dystopian technology — is now @museband. Zuckerberg accidentally tagged the actual band when he announced it.
Nature: What the Navier-Stokes Result Actually Means — Nature's explainer on what the math says, what the Clay Prize requires versus what OpenAI proved, and why Terence Tao praised the human-led work. Less PR, more math.
Your take: Apple went foldable, OpenAI threw 10,000 agents at a math problem, DeepSeek put a 552B open-weight model at $0.30/M that beats GPT-5.6 Sol. Which one actually changes how you work? Drop it in the comments 👇
If you missed our previous updates, don’t worry, here they are: GPT-6 Astra, Claude Fable 5.1 + EFS, OpenAI Drops Cursor | Weekly Digest