Skip to content

Latent Space

AI engineering newsletter and podcast that publishes the daily AINews recap of model releases, papers and community reaction.

Known aliases

  • AINews
  • AI News
  • Latent Space
  • latent.space

Current stories

buildOne report1 publisher

Two Sigma lets classifiers decide routine agent calls and sends people only the important ones

OpenAI, Anthropic and Two Sigma speakers framed human review in financial agents as a placement choice, latent.space reported from AI Engineer New York. Only Two Sigma's speaker is quoted attacking blanket approval; the lab remarks describe gates and design questions.

Reality

Evidence30
Adoption
Not scored
Hype gap+30
Incentives65
Confidence55
buildOne report1 publisher

Dyna chief scientist says its models folded restaurant napkins at 99.4% success over 24 hours

Dyna Robotics says its Dyna-2.1 robot agent folded restaurant napkins at a 99.4% success rate for 24 straight hours. The figure comes from a podcast's summary of a conference talk, with no napkin count or definition of success attached.

Publishers:latent.space

Reality

Evidence30
Adoption35
Hype gap+30
Incentives70
Confidence45
buildOne report1 publisher

Meta's Muse Wonder Full trails Opus 5.5 by two points on the Artificial Analysis index

Meta's Muse Wonder Full scores 56 on Artificial Analysis's Intelligence Index, two points behind Opus 5.5, according to latent.space. Full is limited to Muse Code subscribers, so the Meta Model API serves only Wonder.

Publishers:latent.space

Reality

Evidence58
Adoption
Not scored
Hype gap+12
Incentives
Not scored
Confidence60
buildOne report1 publisher

Pi Durable checkpoints every agent step so crashed runs resume from storage

Pi Durable ports the Pi agent harness to TypeScript and checkpoints every step to one of three stores, so crashed agents resume where they stopped. Whether it can replace hand-built resume code depends on how it treats a tool call cut off mid-flight.

Publishers:latent.space

Reality

Evidence28
Adoption
Not scored
Hype gap+30
Incentives
Not scored
Confidence35
investOne report1 publisher

Bespoke Nimble alone closed 24 of the 27-point gap to Jev within two days of launch

Bespoke Nimble's LoRA fine-tune of Qwen3.5-9B scored 90% against Jev's 93% on an eval it curated itself, and five more replications landed at scales between 421M and 35B parameters. No standard benchmark exists for the category yet.

Reality

Evidence32
Adoption46
Hype gap+42
Incentives72
Confidence38
investOne report1 publisher

Databricks reports 60% higher coding spend on the model benchmarks score as cheaper

A Latent Space roundup says Databricks' spend rose about 60 percent when its AI engineers moved to GPT-6 Astra, a model many benchmarks rank cheaper per task. The rollout covers roughly 3,500 engineers.

Reality

Evidence32
Adoption55
Hype gap+15
Incentives60
Confidence38
buildOne report1 publisher

Databricks reports 60 percent higher coding spend after switching to Astra

Astra is rated cheaper per task on many benchmarks. After putting it in front of roughly 3,500 engineers, Databricks reports about 60 percent more coding spend.

Publishers:latent.space

Reality

Evidence38
Adoption62
Hype gap+30
Incentives60
Confidence45
buildOne report1 publisher

AIUC raises $40M to put an insurance policy behind its AIUC-1 agent audit

Rune Kvist's startup wants to sell enterprises an audited, insured answer to who pays when an agent fails, and the round was announced in a podcast episode that names Cursor, Harvey, Lovable and ElevenLabs.

Publishers:latent.space

Reality

Evidence42
Adoption20
Hype gap+35
Incentives78
Confidence45
productOne report1 publisher

Flue closes external pull requests on arrival and refiles them as issues

Fred Schott's Flue treats an outside patch as a lead to qualify rather than work a maintainer owes review. Vercel's AI SDK team says plainly that it trusts its own tuned agents more than the community's.

Publishers:substack.com

Reality

Evidence58
Adoption66
Hype gap+26
Incentives72
Confidence57
buildOne report1 publisher

Fal reached continuous video generation by tuning inference on someone else's model

AINews reports that fal post-trained Minimax's H3 and served it 35x faster than the official endpoint. Twitch and YouTube removed the resulting live stream immediately, so fal stood up its own player instead.

Publishers:latent.space

Reality

Evidence22
Adoption15
Hype gap+55
Incentives72
Confidence34
buildOne report1 publisher

Ng's new DeepLearning.ai spec is a reskilling checklist, not a hiring req

Built from 10,000-plus job postings and interviews with hiring managers, the four-skill list describes what buyers already want. Three of the four need no machine learning background.

Publishers:latent.space

Reality

Evidence38
Adoption34
Hype gap+22
Incentives74
Confidence41
buildOne report1 publisher

The judge went synthetic first, which tells you which part of your pipeline is next

Latent Space dates five flips from human-made to model-made since 2022. The order they happened in predicts more than the annual cadence the essay claims for it.

Publishers:latent.space

Reality

Evidence42
Adoption58
Hype gap+32
Incentives55
Confidence48
buildOne report1 publisher

Write your agent harness like a lease: most of it has an absorption date

latent.space argues models and harnesses improved together, and that models keep swallowing the harness. If so, most scaffolding you write is scheduled for deletion. Permissions are not.

Publishers:latent.space

Reality

Evidence42
Adoption48
Hype gap+18
Incentives
Not scored
Confidence38
buildOne report1 publisher

Simile's simulated shoppers: two 85 percents, one budget line

A $2B Series B rests on synthetic respondents agreeing with human focus groups 85 to 99 percent of the time. The band's floor and ceiling are not the same product.

Publishers:latent.space

Reality

Evidence24
Adoption28
Hype gap+46
Incentives78
Confidence52
buildOne report1 publisher

Pocock's /wayfinder bets that the bottleneck in overnight agents is planning, not code

The new skill exists because its author found the planning stage, not the execution stage, was the thing rationing how much work he could hand to agents overnight.

Publishers:latent.space

Reality

Evidence30
Adoption16
Hype gap+32
Incentives68
Confidence36
buildOne report1 publisher

Stripe paid $7B for a router. Glean is selling the choice not to call a model at all

Glean says its customers mostly turn on automatic model selection to control spend, not to improve answers. The routing layer, not the model, is where enterprise AI budgets now get decided.

Publishers:latent.space

Reality

Evidence32
Adoption64
Hype gap+34
Incentives82
Confidence44
buildOne report1 publisher

Flue 2 bets that agents are a rendering problem, not an orchestration one

Fred Schott's first stable release of Flue replaces file-based routing with React-style hooks, on the argument that an agent re-renders every turn and the harness defines it.

Publishers:latent.space

Reality

Evidence44
Adoption21
Hype gap+34
Incentives72
Confidence46