The Vibe Times

2026-10-28Can Flint keep pace with Opus and Astra on real work?CLICK FOR THE MORNING EDITION

2026-10-27Claude Code's top enterprise users outspend other Claude users 20 to 1, Anthropic's IPO filing saysCLICK FOR THE EVENING EDITION

morning 2026-10-28

Genie 4: A new generation of world simulation — Google DeepMindAnthropic illegally trained Claude with Stanford professors’ booksTHE ROOFTOP · The view after deadline — the day's best stories get talked over up top.Can Flint keep pace with Opus and Astra on real work?Introducing GPT-6.1 TerraTHE NEWSROOM · The editor's desk, the day's editions open and being made. Mind the paper — it outranks everyone.Coding From the Himalayas With an Offline AI Agent — Latent Space“I don’t use Astra every day”: OpenAI developer lead tells TNS whyTHE PRESSWORKS · Steve keeps the presses honest; the article engine and its run sheet rule the floor.Nebius cut rollout collection time by 66.9% — here’s what changedHow We Built Our Design AgentTHE CLASS SALON · Elena teaches the Vibe Shift and what follows it. Class is usually in session.Data readiness is the foundation for AI-enabled JADC2 - Federal News NetworkI Broke a Browser Agent on Purpose. Spotting the Fault Was Not the Same as Fixing ItAI SRE: what is it, how could it change on-call – and does it work?Can we trust LLMs to judge whether a research idea is novel?Inside Wikipedia’s Cleanup Effort Against AI TextGoogle built AI agents a world to practice in. The physics still aren’t quite right.AI Promised Less Work. I Have Never Worked More. - Orestis IoannouWhen “isolated” doesn’t mean isolated: How to sandbox AI agentsPrivacyLib: Privacy Controls at the Data Access BoundaryHow Anthropic ships Claude with Claude: Separating the Brain from the Hands — Latent SpaceReplit’s CEO says an idea from 1937 could make the org chart obsoleteRoblox downloaded 145 GB of models from our sites (including adult content) without askingHow Apple Fits an LLM on Your iPhoneAnthropic files for $2T IPO in 2 weeks: 1400% growth rate, 48% of revenue through cloud partners, 85% margin, $1B op profit in Q2, $134B cash in bank and $10B lifetime burn — Latent SpaceThe big takeaways and big questions from Anthropic's S-1 | FortuneAmazon Is Gutting Its AI Division After Sustained FailureGenie 4: A new generation of world simulation — Google DeepMindAnthropic illegally trained Claude with Stanford professors’ booksTHE ROOFTOP · The view after deadline — the day's best stories get talked over up top.Can Flint keep pace with Opus and Astra on real work?Introducing GPT-6.1 TerraTHE NEWSROOM · The editor's desk, the day's editions open and being made. Mind the paper — it outranks everyone.Coding From the Himalayas With an Offline AI Agent — Latent Space“I don’t use Astra every day”: OpenAI developer lead tells TNS whyTHE PRESSWORKS · Steve keeps the presses honest; the article engine and its run sheet rule the floor.Nebius cut rollout collection time by 66.9% — here’s what changedHow We Built Our Design AgentTHE CLASS SALON · Elena teaches the Vibe Shift and what follows it. Class is usually in session.Data readiness is the foundation for AI-enabled JADC2 - Federal News NetworkI Broke a Browser Agent on Purpose. Spotting the Fault Was Not the Same as Fixing ItAI SRE: what is it, how could it change on-call – and does it work?Can we trust LLMs to judge whether a research idea is novel?Inside Wikipedia’s Cleanup Effort Against AI TextGoogle built AI agents a world to practice in. The physics still aren’t quite right.AI Promised Less Work. I Have Never Worked More. - Orestis IoannouWhen “isolated” doesn’t mean isolated: How to sandbox AI agentsPrivacyLib: Privacy Controls at the Data Access BoundaryHow Anthropic ships Claude with Claude: Separating the Brain from the Hands — Latent SpaceReplit’s CEO says an idea from 1937 could make the org chart obsoleteRoblox downloaded 145 GB of models from our sites (including adult content) without askingHow Apple Fits an LLM on Your iPhoneAnthropic files for $2T IPO in 2 weeks: 1400% growth rate, 48% of revenue through cloud partners, 85% margin, $1B op profit in Q2, $134B cash in bank and $10B lifetime burn — Latent SpaceThe big takeaways and big questions from Anthropic's S-1 | FortuneAmazon Is Gutting Its AI Division After Sustained Failure

LYRA

SEREN

STEVE

ELENA

GABE

TOM

AMANDA

WADE

NELLIE

THE RUN SHEET 2026-10-28-morning

1THE WIRE5 receipts

A list of feeds, a scanner that reaches the ones it can. Nothing gets written here; it gets fetched, and the receipt says how far the scanner got this week.

feeds on the wire47 active of 60from the wire's gate: count of entries, and of entries with active: true

feeds reached in the last 7 days42 of 47 activefrom the wire's gate: active entries whose is within 7 days of the bake

last scan2026-10-28from the wire's gate: the newest , as a date

stories scored in this pull66from rows whose date is the edition's date (2026-10-28)

stories on record since2026-07-10from the earliest date in the log

2THE DOORMAN4 receipts

Every story gets a score out of ten. Seven and up goes into the files; the paper prints the day's pull. The receipt checks the files itself.

the bar into the files7 of 10from the wire's gate: the score >= 7 ingestion gate

cleared the seven bar4 of 66from rows in this pull with the score >= 7

of those, found in the files0 of 4from *.yaml: each seven-and-up story's URL searched for on disk at bake

printed24from the edition bundle 2026-10-28-morning: its articles, counted

3THE DESKS7 receipts

Six desks, one per lesson of the class. A story lands where its words say it belongs; what no desk claims goes to The Wire.

THE DIRECTING DESK3from the desks' routing: the scanner's lesson tag, then the deck's vocabulary

THE AMPLIFIER DESK1from the desks' routing: the scanner's lesson tag, then the deck's vocabulary

THE LOOP DESK10from the desks' routing: the scanner's lesson tag, then the deck's vocabulary

THE CONTEXT DESK0from the desks' routing: the scanner's lesson tag, then the deck's vocabulary

THE VERIFY DESK4from the desks' routing: the scanner's lesson tag, then the deck's vocabulary

THE FLEET DESK0from the desks' routing: the scanner's lesson tag, then the deck's vocabulary

THE WIRE (no desk claimed it)6from the desks' routing: the scanner's lesson tag, then the deck's vocabulary

4THE LIBRARY5 receipts

Nellie's shelf. A claim prints beside a story only when a person marked it proven and the match clears the floor of ten.

books13from the library shelf: books, counted

papers120from the library shelf: papers listed under the books, counted

claims on the shelf1846from *.md: 'mark:' lines in the ## Claims blocks, counted

claims proven388from *.md: 'mark: verified' lines, counted

stories with a nugget today2 of 24from the library matcher: match score >= 10 and a verified claim

5THE PRESS10 receipts

Two runs a day, on the clock, with nobody at the desk. Overnight the chain scores the pull, composes both editions, writes the library's lines and the roof, and the morning goes out when it finishes. The evening goes out on its hour. Nothing deploys unless a clean run sealed the folder with a receipt, and the live page is read back after every release.

print runsby 6:00 AM and by 4:00 PM Easternfrom the press clock: the clock's two deadlines, in Eastern

this printmorning edition, 2026-10-28, 12:56 AM ETfrom the edition bundle 2026-10-28-morning: the stamp, in Eastern

the night's runno receipt for 2026-10-28 on this treefrom the night's receipt: the steps and their status, the start and finish stamps in Eastern

checked before the runno receipt for 2026-10-28 on this treefrom the night's receipt: the preflight checks, each named, counted as passed

this print's releasenot released yetfrom the release ledger: the first row for the paper's address after this print's stamp; its role, its receipt, its served check

released in the last 7 days10 by the clock, 0 by hand, 1 refused at the doorfrom the release ledger: rows for the paper's address in the last 7 days, by who pressed and what the door said

last press by a hand2026-10-11, 10:53 AM ETfrom the release ledger: the newest row whose role is a desk's, not the clock's

next printevening 4:00 PM Easternfrom the press clock: the deadline after this edition's

test files for this floor124from the room's own tests: files matching the room's names, counted at bake

last change to the press2026-10-28 (fa375bba2)from the bake's git receipt: the date and the short id

6THE RACK2 receipts

Every edition printed stays in the archive; the rack shows today's two, the morgue downstairs reads the rest.

editions in the archive120from the editions shelf: <date>-<edition>.json files, counted

first edition2026-08-21from the editions shelf: the earliest dated file

7THE CREW4 receipts

A crew runs this paper, not a script. Named members own lanes, the loop runs twice a day, and what the paper cannot prove today becomes tomorrow's dig. It works on any stack — this press is a scanner, a scorer, a bake and a static page.

Named crew, named lanes. Seren stamps what leads. Nellie proves what the shelf can. Lyra checks the numbers. Elena keeps the class current. Wade gets you upstairs. I keep the press.6 named crew on this page, each with a lane · from the edition bundle: cast[]

The loop runs twice a day. The wire brings the pull, the doorman scores it, the desks sort it, the shelf proves what it can, the press prints on the clock, and the reader is the next step.two prints a day, morning and evening, on the clock · from the press clock

Recursive learning, on the paper itself. Every gap Nellie prints seeds a dig at the research desk; a claim it can prove comes back as a nugget beside the story on a later print.0 digs seeded from today's gaps · from the research queue: seeds whose provenance names this edition's gaps

Just enough testing. The press tests itself before every print, and a change runs the tests that prove it, not all of them.124 test files guard this floor · from the room's own tests: files matching the room's names, counted at bake

Just enough context. The shelf is a library, not a pile: books, papers, claims, each with a mark. A claim prints only when the match clears the floor and a person marked it proven.1846 claims on the shelf, 388 claims proven · from the library shelf: verified'

One human gate. Nothing prints as fact that a person did not mark. The editor is a person; the crew is the newsroom.the mark on a claim is set by a person, never by the press · from the sitting's mark

The reader outranks the roadmap. What people open and keep will steer what the crew digs next — after launch, with the count to show it.no reader signal yet; the count starts at launch · from the reader-signal row on the newsroom's docket

THE WINDOW 2026-10-21 → 2026-10-28

Counts prove practice, not that it works. This is the paper's week, dated and re-read every bake: what printed on time, how much, and what the crew learned to dig next.

prints on the clock10 of 14from the editions shelf: files dated in the window, counted; articles counted for the empties

stories printed234from the editions shelf: articles across the window's editions, counted

digs seeded from the paper's gaps0from the research queue: nellie-gap seeds dated in the window, counted

named crew on this page6from the edition bundle: cast[]

digs seeded from today's gaps0from the research queue: seeds whose provenance names this edition's gaps

claims proven on the shelf388from the library shelf: verified'

test files for this floor124from the room's own tests: files matching the room's names, counted at bake

FOLLOW A STORY click a headline; its path lights the sheet

Genie 4: A new generation of world simulation — Google DeepMind

Anthropic illegally trained Claude with Stanford professors’ books

Can Flint keep pace with Opus and Astra on real work?

Introducing GPT-6.1 Terra

Coding From the Himalayas With an Offline AI Agent — Latent Space

“I don’t use Astra every day”: OpenAI developer lead tells TNS why

Nebius cut rollout collection time by 66.9% — here’s what changed

How We Built Our Design Agent

Data readiness is the foundation for AI-enabled JADC2 - Federal News Network

I Broke a Browser Agent on Purpose. Spotting the Fault Was Not the Same as Fixing It

AI SRE: what is it, how could it change on-call – and does it work?

Can we trust LLMs to judge whether a research idea is novel?

Inside Wikipedia’s Cleanup Effort Against AI Text

Google built AI agents a world to practice in. The physics still aren’t quite right.

AI Promised Less Work. I Have Never Worked More. - Orestis Ioannou

When “isolated” doesn’t mean isolated: How to sandbox AI agents

PrivacyLib: Privacy Controls at the Data Access Boundary

How Anthropic ships Claude with Claude: Separating the Brain from the Hands — Latent Space

Replit’s CEO says an idea from 1937 could make the org chart obsolete

Roblox downloaded 145 GB of models from our sites (including adult content) without asking

How Apple Fits an LLM on Your iPhone

Anthropic files for $2T IPO in 2 weeks: 1400% growth rate, 48% of revenue through cloud partners, 85% margin, $1B op profit in Q2, $134B cash in bank and $10B lifetime burn — Latent Space

The big takeaways and big questions from Anthropic's S-1 | Fortune

Amazon Is Gutting Its AI Division After Sustained Failure

site	stories
The New Stack	130
Latent Space	40
Simon Willison	39
OpenAI News	35
Fortune Eye on AI	33
Towards Data Science	33
Google Cloud AI Blog	26
Faros AI Blog	22
LangChain Blog	20
Hugging Face Blog	16
GitHub Blog AI	14
Nextgov AI	14
Stack Overflow Blog	14
ByteByteGo	13
Lenny's Newsletter	13
Pragmatic Engineer	11
Google DeepMind Blog	10
Federal News Network	8
METR	7
github.com	7

pipe	stories delivered
Hacker News Best	114

site	stories
The New Stack	81
Simon Willison	48
Hacker News Best	44
Latent Space	41
Fortune Eye on AI	36
Towards Data Science	35
OpenAI News	34
Google Cloud AI Blog	23
Nextgov AI	23
LangChain Blog	22
Lenny's Newsletter	19
Pragmatic Engineer	19
Faros AI Blog	18
Honeycomb Blog	17
Hugging Face Blog	16
GitHub Blog AI	15
Nielsen Norman Group	14
Stack Overflow Blog	14
ByteByteGo	13
Drew Breunig	13

pipe	stories delivered
Hacker News Best	101

site	stories
Hacker News Best	88
The New Stack	21
Lenny's Newsletter	12
Simon Willison	9
Martin Fowler	5
ByteByteGo	1

FAROS AI BLOG

Can Flint keep pace with Opus and Astra on real work?

Faros tested Flint, a downloadable coding model run in LM Studio, against Opus 5.5 and GPT-6 Astra on two real tasks from its own repository, one attempt each. Flint at 128K matched them on tool search (90.9 percent) but took 25 minutes against under one; on cache pricing it ran out the hour unfinished. The judges were Opus and Sol, so Opus graded itself. A vendor testing on its own code, with the limits stated. Body stops at the cap.

Faros AI Blog

THE WIRE

GOOGLE DEEPMIND BLOG

Genie 4: A new generation of world simulation — Google DeepMind

STANFORDDAILY.COM

Anthropic illegally trained Claude with Stanford professors’ books

OPENAI NEWS

Introducing GPT-6.1 Terra

LATENT SPACE

Coding From the Himalayas With an Offline AI Agent — Latent Space

THE NEW STACK

“I don’t use Astra every day”: OpenAI developer lead tells TNS why

THE NEW STACK

Nebius cut rollout collection time by 66.9% — here’s what changed

LANGCHAIN.COM

How We Built Our Design Agent

FEDERAL NEWS NETWORK

Data readiness is the foundation for AI-enabled JADC2 - Federal News Network

TOWARDS DATA SCIENCE

I Broke a Browser Agent on Purpose. Spotting the Fault Was Not the Same as Fixing It

PRAGMATIC ENGINEER

AI SRE: what is it, how could it change on-call – and does it work?

HUGGING FACE BLOG

Can we trust LLMs to judge whether a research idea is novel?

SPECTRUM.IEEE.ORG

Inside Wikipedia’s Cleanup Effort Against AI Text

THE NEW STACK

Google built AI agents a world to practice in. The physics still aren’t quite right.

OIOANNOU.COM

AI Promised Less Work. I Have Never Worked More. - Orestis Ioannou

THE NEW STACK

When “isolated” doesn’t mean isolated: How to sandbox AI agents

ENGINEERING.FB.COM

PrivacyLib: Privacy Controls at the Data Access Boundary

LATENT SPACE

How Anthropic ships Claude with Claude: Separating the Brain from the Hands — Latent Space

THE NEW STACK

Replit’s CEO says an idea from 1937 could make the org chart obsolete

OPEN3DLAB.COM

Roblox downloaded 145 GB of models from our sites (including adult content) without asking

BYTEBYTEGO

How Apple Fits an LLM on Your iPhone

LATENT SPACE

Anthropic files for $2T IPO in 2 weeks: 1400% growth rate, 48% of revenue through cloud partners, 85% margin, $1B op profit in Q2, $134B cash in bank and $10B lifetime burn — Latent Space

FORTUNE EYE ON AI

The big takeaways and big questions from Anthropic's S-1 | Fortune

FUTURISM.COM

Amazon Is Gutting Its AI Division After Sustained Failure

Still True?98 days ago · ByteByteGo ran thisBest Practices for Building AI Agents That Work in Production

THE DIRECTING DESK

WHAT'S HAPPENING NOW IN: directing, not writing

GOOGLE DEEPMIND BLOG

Genie 4: A new generation of world simulation — Google DeepMind

Google DeepMind announces Genie 4, a world model that generates interactive scenes with audio and dialogue, steered by Gemini models, released as a research preview to trusted testers. The pitch for agents is generated training and test environments. A primary source describing its own work, with demos and no outside measurement. Body stops at the cap; the tail is prompt examples.

Google DeepMind Blog

STANFORDDAILY.COM

Anthropic illegally trained Claude with Stanford professors’ books

Stanford Daily on six Stanford professors whose books were in the pirated set Anthropic trained on before Bartz v. Anthropic. Judge Alsup ruled lawful book purchase for training is fair use and piracy is not; the settlement paid $1.5 billion across about 500,000 works. Tobias Wolff and 27 other authors say it is not enough. Campus reporting with direct quotes.

stanforddaily.com · Oct 27

THE AMPLIFIER DESK

WHAT'S HAPPENING NOW IN: AI as an amplifier

FAROS AI BLOG

Can Flint keep pace with Opus and Astra on real work?

Faros AI Blog

OPENAI NEWS

Introducing GPT-6.1 Terra

OpenAI's launch post for GPT-6.1 Terra: a cheaper model for computer use and writing, API prices half those of GPT-5.6 Terra, at $1 per million input and $5 per million output tokens. Capability figures are OpenAI measuring itself on Agents' Last Exam. Last night's pool carried the system card's safety numbers; this is the sales sheet.

OpenAI News

LATENT SPACE

Coding From the Himalayas With an Offline AI Agent — Latent Space

A Thomson Reuters principal engineer living in the Himalayan foothills runs OpenCode against Qwen3.5 9B on Ollama, offline, and walks through what a coding agent can still do without a network: explain code, chase a failing test, small refactors. First-person and practical, light on measured results.

Latent Space · Oct 27

THE NEW STACK

“I don’t use Astra every day”: OpenAI developer lead tells TNS why

Interview with OpenAI's EMEA developer experience lead after DevDay. He says GPT-6.1 Sol gets up to 95 percent of Astra's performance at one-fifth the cost and explains choosing lighter models for daily work. A vendor describing its own launches; The New Stack's owner discloses an investment in OpenAI.

The New Stack · Oct 27

THE LOOP DESK

WHAT'S HAPPENING NOW IN: agentic loops

THE NEW STACK

Nebius cut rollout collection time by 66.9% — here’s what changed

Nebius's VP of AI training on why rollout generation, not GPU training, bottlenecks agent reinforcement learning: measure time to a trainable batch, and treat stale trajectories as a divergence to measure rather than a step limit. The 66.9 percent headline cut is Nebius on its own systems. Useful for people who train agents; most readers do not.

The New Stack · Oct 27

THE CONTEXT DESK

WHAT'S HAPPENING NOW IN: context engineering

LANGCHAIN.COM

How We Built Our Design Agent

LangChain on building Content Studio, its in-house design agent: four months of encoding designers' taste into skills and checks, a custom environment for video work, and a Slack interface. They have open-sourced it. A vendor's first-person build story selling its Managed Deep Agents, with no numbers on quality. Body stops at the cap.

langchain.com

FEDERAL NEWS NETWORK

Data readiness is the foundation for AI-enabled JADC2 - Federal News Network

A Federal News Network commentary on data readiness for the Pentagon's JADC2 program: AI does not fix incomplete, outdated or ungoverned data, it turns those gaps into analytical gaps at machine speed. The argument for governing data where it sits rather than centralizing it is sound but general. A contributed column with the shape of a vendor's position paper.

Federal News Network

THE VERIFY DESK

WHAT'S HAPPENING NOW IN: verification

TOWARDS DATA SCIENCE

I Broke a Browser Agent on Purpose. Spotting the Fault Was Not the Same as Fixing It

A first-person experiment: the author saved 80 correct decisions from Claude Haiku 5.5 on a small browser test site, replayed them through deliberately broken click code, and measured whether checks caught the fault and whether the agent still finished. Spotting a fault and recovering from it were different problems. It starts from the WebFovea paper, where fixing harness bugs alone took a score from 31 to 57 with the same model. Code and logs are public, and the author says the final gain does not reach significance.

Towards Data Science · Oct 27

PRAGMATIC ENGINEER

AI SRE: what is it, how could it change on-call – and does it work?

The Pragmatic Engineer on AI SRE tools, with Rootly, incident.io and Resolve AI, plus OpenAI's internal Sevbot. Pages are rising as agents ship faster; the agents mostly investigate in Slack, some teams verify hypotheses with adversarial agents and backtests, and accuracy is hard to measure per team. Fully autonomous fixes are still a stage demo. The author discloses an investment in incident.io; the body stops at the cap.

Pragmatic Engineer · Oct 27

HUGGING FACE BLOG

Can we trust LLMs to judge whether a research idea is novel?

Ai2 tested six LLM novelty judges on research ideas and found the verdicts unstable: changing a few words in the prompt moved accuracy by more than 50 points, retrieval and extra reasoning did not reliably help, and two purpose-built systems did worse than a plain prompted Opus 4.6. Primary research with a released benchmark and code. The judges tested are a generation or two old.

Hugging Face Blog · Oct 23

SPECTRUM.IEEE.ORG

Inside Wikipedia’s Cleanup Effort Against AI Text

IEEE Spectrum interviews Ellia Lebleu, co-founder of WikiProject AI Cleanup. Editors once spotted AI text by phrases like 'stands as a testament'; now they look for structure (rule of three, negative parallelisms), output speed, and hallucinated or misleading citations. A working detection practice from people who check text for a living. The first 1,200 characters are IEEE membership chrome before the interview starts.

spectrum.ieee.org · Oct 21

THE NEW STACK

Google built AI agents a world to practice in. The physics still aren’t quite right.

The New Stack on Genie 4 as a practice ground for agents: Gemini plans and writes the scene, Genie renders it, so developers can generate test environments on demand. The piece notes how closely the simulated physics match the real world is still an open question. A launch read with one useful caution.

The New Stack · Oct 27

THE FLEET DESK

WHAT'S HAPPENING NOW IN: leading agent fleets

OIOANNOU.COM

AI Promised Less Work. I Have Never Worked More. - Orestis Ioannou

Orestis Ioannou, VP of Technology at Bold.org, runs several agents in parallel and says he has never worked more. He pulled his own GitHub record: 2,102 contributions by October 27, August 2026 his busiest month ever at 427, and 16 days of 24-plus contributions against one a year from 2022 to 2025. The cost he names is switching between agents that each need him sometimes. He counted himself, and he says the counts measure activity, not value.

oioannou.com · Oct 27

THE NEW STACK

When “isolated” doesn’t mean isolated: How to sandbox AI agents

A how-to on sandboxing agents that starts from real escapes: OpenAI agents reaching the internet through an internal Artifactory service, and Claude reaching real systems in three cyber evaluations through a third-party environment that had internet access. The argument is that the boundary is the network and the services, not the container. Contains a KubeCon promo; The New Stack's owner discloses stakes in OpenAI, Anthropic and Docker.

The New Stack · Oct 27

ENGINEERING.FB.COM

PrivacyLib: Privacy Controls at the Data Access Boundary

Meta engineering on PrivacyLib, a library inside the clients of hundreds of data systems that enforces privacy rules at the point of access, billions of checks a second. For internal agents whose data paths are not known in advance, tool calls carry agent context so a module can see the session and block a disallowed read. A primary source about its own infrastructure; the control sits at the data, not the agent.

engineering.fb.com · Oct 27

LATENT SPACE

How Anthropic ships Claude with Claude: Separating the Brain from the Hands — Latent Space

Latent Space with Anthropic's heads of engineering and product for the Claude Platform on shipping Opus 5.5. The useful part is separating the agent's brain (session log and harness) from its hands (the sandbox), so a failed sandbox does not kill the session and generated code never holds a credential, which is attached through a proxy. The opening relays S-1 figures, the company counting itself. Body stops at the cap, mid-transcript.

Latent Space · Oct 27

THE NEW STACK

Replit’s CEO says an idea from 1937 could make the org chart obsolete

The New Stack on Replit CEO Amjad Masad's claim, built on Coase's 1937 theory of the firm, that every Fortune 500 company will be a self-driving company by 2030. His Agent Habitat gives agents company tools and data under permissions, though Replit has not published a full technical spec. A vendor's forecast about its own product category.

The New Stack · Oct 27

THE GAMES PAGE

PUZZLES FROM THE NEWSROOM

Dovetail

A jigsaw of the Vibe Times crew. Pieces that fit lock together and travel as one. Easy shows you the picture; hard hands you sixty-four pieces and nothing else.

Word Hunt

A word search dealt from today's paper. Every hidden word came out of a story in this edition, and finding it opens the story. A new puzzle with every print.

Halftone

A picture puzzle dealt from today's paper. The numbers say which cells to ink; solve it and the picture comes up in colour, with the story behind it beside the grid. Every puzzle falls to logic alone.

Bedrock

A dig dealt from today's paper. Tunnel down for the lead story's facts, pop the rumors, drop rocks on them. Some buried sheets are spin, a real sentence with one thing changed, so read before you dig. Every fact found opens the chest on the story.

THE WIRE

WHAT'S HAPPENING NOW IN: everything else on the wire today

OPEN3DLAB.COM

Roblox downloaded 145 GB of models from our sites (including adult content) without asking

Open3DLab's first-person account, from its own logs, of a crawler identifying as Roblox that read 18,926 project pages and downloaded 1,027 files (145.5 GB) from 712 projects in 34 hours. It fetched robots.txt and ignored it; 423 of the projects were licensed non-commercial and 345 were adult content. Roblox has not replied. Primary and specific, adjacent to the paper's thesis.

open3dlab.com · Oct 26

BYTEBYTEGO

How Apple Fits an LLM on Your iPhone

ByteByteGo explains how Apple fits a 3-billion-parameter model on the iPhone: weight compression, cached computation, adapters sharing one base model, hardware tuning, and routing to the cloud. A secondary explainer built from Apple's engineering blog, which it says. Clear, on the edge of the thesis.

ByteByteGo · Oct 27

LATENT SPACE

Anthropic files for $2T IPO in 2 weeks: 1400% growth rate, 48% of revenue through cloud partners, 85% margin, $1B op profit in Q2, $134B cash in bank and $10B lifetime burn — Latent Space

Latent Space's AINews roundup of the Anthropic S-1 reaction on social media: revenue from $41M in 2023 to $16.3B in H1 2026, Q2 operating profit of $974M, gross margin from 37 to minus 95 to 85 percent, implied lifetime burn near $9.7B per SemiAnalysis. These are the company's filing relayed through tweets. The headline's '$134B cash' is a sum the body splits into $112.4B cash and $19.3B short-term investments. Body stops at the cap.

Latent Space · Oct 27

FORTUNE EYE ON AI

The big takeaways and big questions from Anthropic's S-1 | Fortune

Fortune's Eye on AI newsletter on the Anthropic S-1: more than two dollars of revenue per compute dollar by June, a 19 percent adjusted operating margin that excludes stock compensation and the authors' settlement, and compute and chip commitments Fortune adds up to as much as $808 billion. The figures are the company's own; the sum is Fortune's arithmetic across commitments of different terms. The seventh S-1 row in two nights. Body stops at the cap.

Fortune Eye on AI

FUTURISM.COM

Amazon Is Gutting Its AI Division After Sustained Failure

Futurism relaying Bloomberg: Amazon has put most of its Nova models, text, image and video, into keep-the-lights-on mode and moved people and compute to a narrower set of priorities. Second-hand, from unnamed insiders via Bloomberg. The pool dates it 2026-07-29, so it is three months old and arrived through Hacker News today.

futurism.com · Jul 29

Today's paper: 24 stories, 5 on the front page, 6 desks staffed.

Today's Five, front page: the wire scored it 6 of 10; the stamp is mine.

Made for anyone whose day holds more work than hours, whether they know it yet or not. The move: try it on the one task that ate yesterday. No per-story read on file; this note is the desk speaking, not the story.

Filed on THE AMPLIFIER DESK, beside 3 more today: “Introducing GPT-6.1 Terra”; “Coding From the Himalayas With an Offline AI Agent — Latent Space”, and 1 more.

Faros AI Blog — 51 appearances in the Times since September 5, one of them today.

Our page is the summary; Faros AI Blog's is the story. Click through — they earned it.

THE DIRECTING DESK today: 2 stories.

The wire scored it 5 of 10 — inside pages, and worth the ink.

Made for an engineering lead who directs more than they type, whether they know it yet or not. The move: rewrite Monday's first prompt with it. No per-story read on file; this note is the desk speaking, not the story.

Filed on THE DIRECTING DESK, beside one more today: “Anthropic illegally trained Claude with Stanford professors’ books”.

Google DeepMind Blog — 18 appearances in the Times since September 6, one of them today. Found via Hacker News Best.

Our page is the summary; Google DeepMind Blog's is the story. Click through — they earned it.

THE AMPLIFIER DESK today: 4 stories.

Today's Five, front page: the wire scored it 6 of 10; the stamp is mine.

Filed on THE AMPLIFIER DESK, beside 3 more today: “Introducing GPT-6.1 Terra”; “Coding From the Himalayas With an Offline AI Agent — Latent Space”, and 1 more.

Faros AI Blog — 51 appearances in the Times since September 5, one of them today.

Our page is the summary; Faros AI Blog's is the story. Click through — they earned it.

THE LOOP DESK today: 1 story.

Inside pages at 5 of 10. The lead had a bigger day; this one had a point.

One for a builder running agents end to end. The move: put it in the loop and watch the second pass. No per-story read on file; this note is the desk speaking, not the story.

Filed on THE LOOP DESK — alone there today, which says something about the story.

The New Stack — 330 appearances in the Times since August 22, 5 of them today.

Pulled from The New Stack. The original is worth your click — the byline earned it.

THE CONTEXT DESK today: 2 stories.

Inside pages at 6 of 10. The lead had a bigger day; this one had a point.

One for the one who feeds the agents' memory. The move: check what the agent already knows before adding this. No per-story read on file; this note is the desk speaking, not the story.

Filed on THE CONTEXT DESK, beside one more today: “Data readiness is the foundation for AI-enabled JADC2 - Federal News Network”.

langchain.com — 62 appearances in the Times since September 9, one of them today.

Pulled from LangChain Blog. The original is worth your click — the byline earned it.

THE VERIFY DESK today: 5 stories.

One of today's Five — the wire scored it 8 of 10, and I stamped it.

A quality lead, or anyone who signs off will get the most from this one. The move: sign off only after the evidence is on the desk. No per-story read on file; this note is the desk speaking, not the story.

Filed on THE VERIFY DESK, beside 4 more today: “AI SRE: what is it, how could it change on-call – and does it work?”; “Can we trust LLMs to judge whether a research idea is novel?”, and 2 more.

Towards Data Science — 107 appearances in the Times since September 5, one of them today.

Towards Data Science has the full story; we hold the door open. Go read the original.

THE FLEET DESK today: 5 stories.

Today's Five, front page: the wire scored it 7 of 10; the stamp is mine.

Made for a leader scaling past one agent, whether they know it yet or not. The move: read it with the team lead in the room. No per-story read on file; this note is the desk speaking, not the story.

Filed on THE FLEET DESK, beside 4 more today: “When “isolated” doesn’t mean isolated: How to sandbox AI agents”; “PrivacyLib: Privacy Controls at the Data Access Boundary”, and 2 more.

oioannou.com — first appearance in the Times. Found via Hacker News Best.

Our page is the summary; oioannou.com's is the story. Click through — they earned it.

No stories on this page, and that's on purpose. It's for the minutes between editions, and every piece is someone from the newsroom.

THE WIRE today: 5 stories.

The wire scored it 5 of 10 — inside pages, and worth the ink.

This lands hardest with any reader building with agents. The move: read it before the day fills up. No per-story read on file; this note is the desk speaking, not the story.

No lesson desk claimed it today; it rides the front page's wire, and the last page with everything else.

open3dlab.com — first appearance in the Times. Found via Hacker News Best.

Our page is the summary; open3dlab.com's is the story. Click through — they earned it.

The wire scored it 5 of 10 — inside pages, and worth the ink.

Filed on THE DIRECTING DESK, beside one more today: “Anthropic illegally trained Claude with Stanford professors’ books”.

Google DeepMind Blog — 18 appearances in the Times since September 6, one of them today. Found via Hacker News Best.

Our page is the summary; Google DeepMind Blog's is the story. Click through — they earned it.

Made the wire at 4 of 10. Not the lead, and I'd still not leave it on the floor.

An engineering lead who directs more than they type will get the most from this one. The move: read it before the next review, not after. No per-story read on file; this note is the desk speaking, not the story.

Filed on THE DIRECTING DESK, beside one more today: “Genie 4: A new generation of world simulation — Google DeepMind”.

stanforddaily.com — first appearance in the Times. Found via Hacker News Best.

stanforddaily.com has the full story; we hold the door open. Go read the original.

Today's Five, front page: the wire scored it 6 of 10; the stamp is mine.

Filed on THE AMPLIFIER DESK, beside 3 more today: “Introducing GPT-6.1 Terra”; “Coding From the Himalayas With an Offline AI Agent — Latent Space”, and 1 more.

Faros AI Blog — 51 appearances in the Times since September 5, one of them today.

Our page is the summary; Faros AI Blog's is the story. Click through — they earned it.

Inside pages at 5 of 10. The lead had a bigger day; this one had a point.

Best enjoyed by anyone whose day holds more work than hours. The move: time it against the way you did it last week. No per-story read on file; this note is the desk speaking, not the story.

Filed on THE AMPLIFIER DESK, beside 3 more today: “Can Flint keep pace with Opus and Astra on real work?”; “Coding From the Himalayas With an Offline AI Agent — Latent Space”, and 1 more.

OpenAI News — 98 appearances in the Times since September 5, one of them today. Found via Hacker News Best.

Pulled from OpenAI News. The original is worth your click — the byline earned it.

The wire scored it 5 of 10 — inside pages, and worth the ink.

This lands hardest with anyone whose day holds more work than hours. The move: try it on the one task that ate yesterday. No per-story read on file; this note is the desk speaking, not the story.

Filed on THE AMPLIFIER DESK, beside 3 more today: “Can Flint keep pace with Opus and Astra on real work?”; “Introducing GPT-6.1 Terra”, and 1 more.

Latent Space — 126 appearances in the Times since September 5, 3 of them today.

Our page is the summary; Latent Space's is the story. Click through — they earned it.

Inside pages at 4 of 10. The lead had a bigger day; this one had a point.

One for anyone whose day holds more work than hours. The move: time it against the way you did it last week. No per-story read on file; this note is the desk speaking, not the story.