The AI story is leaving brute force behind. Models that think deeper without growing bigger, software teams where the bottleneck is no longer the code, and a searchable map of every single-letter change in the human genome. The golden age of building is here; the scarce skill is knowing what to build.

Key Signals from the Transmission

  • Architecture over size. Looped (recursive) transformers run the same block of layers several times instead of stacking more of them. Mixture-of-Recursions (NeurIPS 2025, KAIST and Google DeepMind) adds a router that decides, token by token, how many passes each one gets: answer the easy part instantly, loop on the hard part. At 135M to 1.7B parameters it beat vanilla models at equal compute with up to 2.18x higher throughput. Deeper thinking without a bigger model.
  • The bottleneck moved. Coding agents made raw code cheap, so the constraint slid downstream to product definition, go-to-market and legal. Anthropic told its growth team to hire more product managers, not fewer; Andrew Ng cites a team proposing two PMs per engineer. The old 8:1 engineer-to-PM ratio is compressing toward 1:1 as engineers take product calls and PMs learn to build.
  • Engineering discipline is not dead; typing is. What survives is system architecture, code safety and what the analysis calls vision engineering: deciding what should exist and holding the whole system in your head while agents fill in the parts.
  • Mapping the genome. Google DeepMind's AlphaGenome Atlas (September 8) precomputes the predicted molecular effect of roughly 9 billion single-letter DNA variants, every possible one, plus 100 million observed indels: a one-petabyte, searchable simulation space, free for non-commercial research. The caveat from Scientific American: 30 times larger than the AlphaFold database, and far less accurate. It prioritizes what to test; it does not replace the wet lab.
  • From atlas to trial. Insilico dosed the first patient on September 10 in GENESIS-IPF-3, the first Phase 3 of a drug whose target and molecule were both found by generative AI: rentosertib, 320 patients, 52 weeks, 47 centers. Its Phase 2a data showed six proteomic aging clocks running 3 to 4 years backwards. The drug is investigational and approved nowhere; the aging signal cannot yet be separated from the disease effect.
  • The next security frontier. Beyond autonomous code exploits: real-time voice filters that clone a familiar voice on a live call from a few seconds of audio. Defenses that actually work are analog: call back on a known number, agree a family password, never move money on a voice alone.

What the Independent Numbers Say

  • Mixture-of-Recursions (Bae et al., NeurIPS 2025): a Recursive Transformer reuses a shared stack of layers across recursion steps; lightweight routers assign each token its own recursion depth; at 135M–1.7B parameters it lowers perplexity and improves few-shot accuracy at equal training FLOPs, with up to 2.18x higher inference throughput. [arXiv]
  • VentureBeat: Anthropic told its growth team to hire more product managers because Claude Code let engineering ship at roughly three times headcount; the traditional 1:8 PM-to-engineer ratio now plays out closer to an effective 1:20. Forbes: Andrew Ng cited a team proposing one PM per 0.5 engineers. [VentureBeat]
  • Google DeepMind, September 8, 2026: AlphaGenome Atlas contains predictions for the effects of 9 billion single-nucleotide variants in the human genome; one petabyte; free for academic research via portal and API. Nature: also covers 100M+ observed indels. Scientific American: 30x larger than the AlphaFold database, far less accurate; research use, not clinically validated. [Google DeepMind]
  • Insilico Medicine, September 10, 2026: first patient dosed in GENESIS-IPF-3 (NCT07687459), a 52-week, 320-patient, 47-center Phase 3 of rentosertib in idiopathic pulmonary fibrosis; the drug remains investigational and unapproved. Phase 2a proteomic data (Nature Biotechnology) showed lower predicted biological age on six aging clocks, roughly 3–4 years. [EurekAlert]
  • Forbes Tech Council: as AI accelerates engineering, the bottleneck migrates upstream to defining the right problems; the PM-to-developer ratio is inverting because strategic product work has not been automated the same way. [Forbes]
Every one of these shifts moves value away from doing and toward deciding. The loop decides how long to think. The atlas decides what to test. The 1:1 team decides what to build. If your role is execution of a spec someone else wrote, the machine is already behind you. Move to the gate where the decision is made, and bring the numbers.

Source: youtube.com/watch?v=ghscjwkkOUY · Baris & Baris, “AI DNA'yı Çözdü, Yaşlanmayı Geri Çeviriyor! Peki Sıradaki Ne?” (Turkish) — curated by SKYTH.