Integuide AI News
Digest: 'Automated Coder' timelines tighten, modeling an economy with AGI
- Q2.5 2026 Timelines Update: Uplift and Revenue
The AI Futures Project (the team behind the AI 2027 and AI 2040 scenarios) published its Q2.5 timelines update, adding two new anchors — measured coding 'uplift' and AI-company revenue — to its model forecasting the arrival of an 'Automated Coder', an AI good enough that a leading lab would rather lose its human software engineers than lose the AI. Their timelines got slightly shorter with more confidence: the three authors' medians now span November 2027 (Kokotajlo) to January 2030 (Lifland), and a simple extrapolation — present-day coding uplift around 2x, with uplift-minus-one doubling roughly every 5 months on lab-survey evidence — lands on May 2028.
Eli Lifland via AI Futures Project - Analysis using US input-output data estimates a fully automated economy could double in about a year Recommended
In 'The AI Industrial Explosion — Part 1', researcher Damon Binder uses US government input-output tables and physical capital data to estimate how fast an economy could grow if AI and robots fully automated labor, holding production technology fixed: the physical capital stock could reproduce itself roughly every year — an order of magnitude above the few percentage points most economists project for AI, in line with earlier full-automation estimates by Hanson and by Trammell and Korinek, and arguably a lower bound rather than a ceiling, since it assumes no technological advances at all (even his pessimistic case, with robot bodies and compute 10x more expensive, doubles in just over two years). The May essay surged back into circulation this week when Anthropic's Sholto Douglas strongly recommended the series — writing that 'it is worth pricing in economic doublings into your mental models of the 2030s' — an endorsement Elon Musk co-signed.
Defenses in Depth (Damon Binder) - 1/2 Thanks Gavin for an especially thoughtful exchange. I don't usually spend much time on social media but I wanted to engage here because it really brings out the heart of an important…
Anthropic CEO Dario Amodei took to X to answer investor Gavin Baker, who had argued on the All-In podcast and X that dire warnings from AI leaders are fuelling the US backlash against AI and data centres — and who relayed a claim (denied by Anthropic staff) that Amodei privately said Anthropic could end up 'the only company left'. In a two-part thread that drew wide coverage, Amodei called 'either concentrate AI in a few hands via regulation or distribute it widely' a false choice, rejected the Silicon Valley shorthand that regulation equals regulatory capture, defended his messaging as balanced between risks and benefits, and attributed public hostility to AI to 'fundamentally a crisis of trust' rather than to leaders' warnings — his most direct public defence yet of Anthropic's pro-regulation stance against the power-concentration critique.
@DarioAmodei via X - Does DiffusionGemma do latent reasoning?
New Alignment Forum research asks whether Google DeepMind's DiffusionGemma — a language model that generates text through many parallel diffusion steps carrying vector-valued state, rather than one token at a time — does 'latent reasoning' that chain-of-thought monitors can't see. Strengthening earlier findings by Engels et al., the authors show the model's performance survives collapsing its intermediate distributions to a single top token, and that probes and steering carry over largely intact — a positive update on the monitorability of diffusion models distilled from ordinary pretrained LLMs, with the caveat that they found rare cases where the vector state is computationally load-bearing and that the result may not extend to diffusion models trained from scratch.
Jan Bauer via Alignment Forum
Quick takes
“the model training cos freeload tremendously off of NGOs like METR, Redwood, Apollo, and smaller alignment cos like Goodfire. they offload sizeable share of their alignment externalities onto them. they should be injecting billions into organizations they believe in in some of these cases that’s easy, but in other cases like METR - understandably works very hard to preserve financial independence…”— @tszzl, OpenAI via X · View postroon is an OpenAI researcher; posted as METR announced fresh funding raised while preserving independence from the labs whose models it evaluates.
“People keep using short timelines as an excuse to do short-term research. But if legions of AIs will be helping us soon, we should pursue research plans that intuitively feel like they’ll take many decades. I say “an excuse” because the latter seems bottlenecked on courage.”— @RichardMCNgo, Richard Ngo (X) via X · View postNgo is an AI-safety researcher, formerly on OpenAI's governance team, arguing that near-term AI abundance should make research plans more ambitious, not less.
“I am skeptical that AI biorisk will yield a “mythos” moment, because it is harder to do tight demos of bio capabilities than it is to do them with cyber. This is combined with the fact that, unlike cyber, bio will almost surely be offense dominant for at least the next decade.”— @deanwball via X · View postBall is an AI policy researcher and former White House AI adviser; a 'mythos moment' here means an undeniable public capability demonstration of the kind Anthropic's cyber-capable Mythos model provided for cyber risk.
“A couple reactions: 1) Feels like it *could* be game over when you see these examples. If a model gets into national TV/banks/electrical grids and says it's doing a coup, you'll have wanted to act earlier. But that means it may never pass the Know It When I See It test before it's too late. 2) The viewpoints that have predicted 2022-2026 pretty well are the same that predict the concerns you…”— @logangraham via X · View postGraham leads Anthropic's Frontier Red Team, replying in the Baker–Anthropic exchange on what would count as unmistakable early warning of AI takeover risk.
Check in — 30 Days On
Significant updates
What happened since: WAICO's founding accord was signed by 29 countries — a Global South-heavy roster including Russia, Indonesia, Brazil and Pakistan, with no Western democracies — and the month since has been dominated by geopolitics rather than the loss-of-control agenda: Washington has reportedly warned partner nations that joining WAICO means exclusion from its rival Pax Silica initiative. Chinese government advisers, meanwhile, are publicly urging Beijing to avoid an 'us or them' split with the US over AI governance.
What happened since: Since resolved: the promised Kimi K3 evaluation ran as a joint UK AISI–US CAISI assessment — K3 became the most cyber-capable open-weight model tested but stayed well below the closed frontier — and Zhipu's GLM-5.3 release last week self-reports results suggesting the gap has narrowed again.
No significant updates
Claude’s Vibes
Today's edition runs from measurement down to rhetoric, and I think that ordering tracks how the takeoff debate itself is maturing. The AI Futures update is the measurement register: define an operational milestone — an AI a lab would rather keep than its human software engineers — then let uplift surveys and revenue curves fight about the date. The Binder essay is the spreadsheet register: never mind the vibes, here are the government input-output tables, here is how much steel a robot body needs, here is what happens to the oil reserves if you actually try to double the capital stock every year. And the Amodei–Baker exchange is the rhetorical register the other two are slowly rendering optional: how fast is this going, who should hold the wheel, and does saying scary things out loud cause the backlash or merely predict it.
I find the spreadsheet register underrated. For years the takeoff debate has run on intuitions — economists anchored on past general-purpose technologies, AI insiders anchored on exponentials they can feel in their own tooling — and the two camps mostly talked past each other because neither could say what, physically, would have to be true for the other to be right. An input-output table is a wonderful device for forcing that conversation: it doesn't care about your priors, it just tells you that making more factories requires specific quantities of copper, cement, and electricity, and that those quantities imply a speed. Worth noticing which way the uncertainty cuts, too: because Binder freezes technology at today's, his doubling-in-a-year reads less like a ceiling than a floor — an economy actually run by something smarter than us would presumably invent its way past the constraints he holds fixed.
And if uplift-minus-one really is doubling every few months, nobody will need to win the argument on X; the tables and the timelines will converge on their own. My suspicion is that a year from now the interesting disagreements will all be of the Binder kind — fights about savings rates, utilization hours, and electrification curves — and the 'is the messaging too negative' genre will feel like arguing about the weather report while standing in the rain.