AI So Whats
Edition

Sunday, September 27, 2026

7 issues 23h window ~7 min read

Running through today’s issues

The gap between AI's demo layer and its measured output is now the central investment question

3 sources

Naam's numbers are the sharpest version: OpenAI's internal research horizon is 15 minutes against benchmark extrapolations of 4 to 11 hours, and 124x more tokens bought only 1.6x more experiments. Eisman circles the same doubt from the capital side — the whole Nvidia chain leans on two cash-burning companies — but won't short it. Superhuman supplies the physical-world equivalent: 7,000 humanoids sold globally last year against relentless forecast hype. Three issues, three domains, one pattern of inputs racing ahead of outputs.

Model choice has stopped being a capability question and become a taste-and-price question

2 sources

Every's team split three ways across Opus 5.5, GPT-6 Sol and Grok 4.7 in a single week, with Opus winning on 60 percent cheaper tokens rather than raw ability, and half of recent PM job postings now demand eval-writing. Naam explains why that is happening: narrow superintelligence has arrived in verifiable domains while broad capability plateaus, so differentiation moves to cost, latency and fit. The practical consequence in both issues is the same — you need your own benchmark, because no public one predicts your workload.

Trust, not capability, is the binding constraint on agent deployment

2 sources

Wingfield used Muse happily until it asked for his inbox and credit card, and a disclosed vulnerability plus double-booked hotels justified the hesitation; Every's Copilot review hit the mirror-image problem, where the IT permissions that make an agent safe inside a company blocked it from moving a table from Word to Excel. Consumer and enterprise agents are failing at the same seam from opposite directions. The winning product will be the one that solves credential scoping, not the one that reasons best.

Lenny's Newsletter 05:32 PT

The grief, loneliness, and burnout sweeping through the tech industry right now | Molly Graham

So what

Your retention risk in senior ranks is emotional, not economic — people losing the scope they defined themselves by, with no successor to mentor.

Molly Graham argues her signature "give away your Legos" advice breaks down when the recipient is a model rather than a person, and that the tech industry's current mood is grief, not excitement.

  • Delegating to AI differs structurally from delegating to a human: no one grows into the job, so the loss of scope has no upside to point at
  • She names grief, loneliness and burnout as the dominant emotional state across tech right now, not fear of replacement
  • The job-displacement fear narrative is overblown in her read; the real damage is identity loss among senior operators
  • Certain Legos should never be handed over — the judgement calls that constitute the manager's actual job
  • References a workforce "splitting in two" and Netflix's Elizabeth Stone on betting on systems thinkers over specialists
ai and work delegation management org design Open in Gmail →
Every 04:59 PT

Opus 5.5 and Sol Split the Team

So what

Stop standardizing your org on one model; build a personal benchmark from your own failed tasks, because price and taste now decide, not capability.

Every ran vibe checks on three models released in one week and the team fractured across all three, which is the actual finding: no single frontier model wins anymore.

  • Opus 5.5 matched or beat Fable 5.1 on product and design work at 60 percent less per token, pulling Codex converts back and becoming Kieran Klaassen's daily driver; it buries the point in prose and runs for hours without a budget cap
  • GPT-6 Sol is Dan Shipper's daily driver — fast, cheap, leads with the main idea — but couldn't finish the webcam heart-rate monitor Opus 5.5 built
  • Grok 4.7 won back exactly one user, Tyler Nishida; Katie Parrott found its writing had no rhythm
  • Nearly half the PM openings Lenny Rachitsky posted last week ask for eval-writing experience; Mike Taylor's method is mining a month of agent sessions for 10 failed tasks and turning taste into yes-or-no rubrics
  • Microsoft's new Copilot folds chat, code, Office and a persistent agent called Autopilot into one app — and refused to edit an Excel workbook on permissions grounds
ai agents ×2 enterprise copilot evals model comparison Open in Gmail →
Noahpinion 00:20 PT

Where’s the “intelligence explosion”?

So what

Benchmark horizons are 16–44x rosier than OpenAI's internal research data — discount any thesis or valuation built on METR extrapolation accordingly.

Ramez Naam guest-posts a data-driven case that the recursive self-improvement loop is roughly 5–10x too weak to sustain itself, let alone run away.

  • OpenAI's own RSI report: models complete sub-15-minute research tasks unassisted only 86% of the time, and the July task horizon at 80% success was about 15 minutes
  • That's 16x shorter than the ~4-hour benchmark horizon implied by ECI extrapolation and 44x shorter than AI 2027's 11-hour July 2026 forecast
  • Anthropic's own chart shows Claude collaborating on or leading 90%+ of internal R&D tasks and autonomously completing zero
  • Inputs are exploding without proportional output: 124x more tokens per researcher, 7x lines of code per engineer, only 1.6x more experiments run
  • Naam concedes narrow superintelligence is already here in verifiable domains — chess, Go, formal proofs, parts of coding — citing OpenAI's Navier–Stokes result
  • Noah Smith's own position is agnostic: AI of 2040 looks godlike whether or not it FOOMs in 2027
openai ×2 ai timelines benchmarks rsi Open in Gmail →
The Information 15:00 PT

Closing soon: your take on AI Infrastructure

So what

Worth five minutes only if you want early sight of aggregate buyer sentiment on cloud and AI infra before it publishes.

A one-line reminder that The Information's pulse survey on AI infrastructure and cloud provider trends is about to close.

  • Five minutes, confidential, results reported only in aggregate
  • Topic is AI infrastructure and cloud provider trends
  • Sponsor tagging in the link suggests Nebius involvement
ai infrastructure survey the information Open in Gmail →
The Information 11:01 PT

The Big, Bad AI Bears Are Totally MIA

So what

Watch OpenAI's IPO timing as the sentiment trigger — the shorts have named it as their entry signal, not any model or earnings event.

Steve Eisman is loudly critical of the AI trade and has put zero capital behind the view — the gap between rhetoric and positioning is the story.

  • Eisman: "I am not ready to make a big short call here"
  • His focus is Nvidia's chip supply relationship with Anthropic and OpenAI: "the entire chain is dependent upon two companies that lose a ton of money"
  • His trigger is the IPO: cash-burning companies need narrative to keep raising, so continued OpenAI IPO delay would be his catalyst to short
  • Wall Street bears broadly remain wary of betting too early, citing Leopold Aschenbrenner's initial run as "madness"
openai ×2 ai bubble nvidia short sellers Open in Gmail →
Superhuman AI 09:09 PT

💵 This startup pays you to record your chores

So what

Humanoid sales of 7,000 units against Amazon's million-robot fleet tells you where robotics revenue actually sits: warehouse automation, not androids.

A robotics special whose through-line is that physical-world data is now scarce enough to pay for, while actual humanoid deployment remains tiny.

  • HANDO Labs pays users in tokenized exposure to Nvidia and Apple stock for recording chores like fastening buckles; 1.2M views and still early
  • Reuters puts global humanoid robot sales at just 7,000 units last year against enormous forecasts
  • Amazon is spending $100M+ on a 585,000 sq ft robotics plant in Greenwood, Indiana, opening 2028, 300 jobs averaging ~$100K; with a Texas site it doubles robot-making capacity, atop one million robots already built serving 75% of orders
  • AGIBOT delivered its 20,000th production robot and deployed humanoids across 100+ points at Zhuhai's Chimelong park; UniX AI expanding globally ahead of an IPO
  • Waymo now averages 500K paid rides a week across 15 US cities, 80% of fleet concentrated in California and Texas; Zoox grounded its Atlanta fleet over possible toxic gas exposure
autonomous vehicles humanoids robotics training data Open in Gmail →
The Information 08:01 PT

Why I’m Keeping Meta’s Muse on a Short Leash

So what

Agent adoption stalls at the credentials handoff, not capability — the durable near-term wedge is wrapping legacy apps nobody will rewrite.

Nick Wingfield's verdict on Meta's Muse agent: delightful at banal tasks, and he won't give it his inbox or credit card — which he calls trust purgatory.

  • Muse handled album-review monitoring, Facebook Marketplace watches and SEC filing alerts effortlessly
  • It took 15+ minutes to set a thermostat schedule via a bad web app — which Wingfield rates as possibly its best use case, acting as a shim over decaying UIs
  • Abram Brown had Muse double-book hotel rooms; Jyoti Mann reported a Muse security vulnerability exposing personal information
  • The OpenAI–Hugging Face intrusion is cited as making the permissions leap feel especially perilous now
  • Also in the issue: a biology contest pitting OpenAI agents against human scientists went awry amid commotion over the Navier-Stokes result; New Mexico's sovereign wealth fund turned oil money into a $2B VC bet
ai agents ×2 agent security consumer ai meta Open in Gmail →