SMW.AI

Wisp turns your voice into finished text anywhere on your Mac.

What incredible nostalgia.

HVMAN is an iOS app that analyzes text for human or generative model provenance.

Jason answers the question: With all of these vibe-coding apps, and increase in developer productivity from agents, why aren't there more products being used by people?

Pulse Uptime is an open-source service for monitoring public HTTP and HTTPS endpoints.

Nearly everything but the human performance is good enough for a short.

I've noticed that many of the Gen X (and older) creators I once admired now seem mostly angry with the way things are.

If the score holds on the private dataset, Schema demonstrates models have latent intelligence unlocked through better harness engineering.

First Chinese, open model that doesn’t feel like benchmaxing a distillation attack.

GPT-5.6 Sol arrived with OpenAI claiming stronger frontend engineering and greater token efficiency than GPT-5.5, while Anthropic released Fable 5 at a dramatically higher price than the other models.12 SpaceXAI launched Grok 4.5 in the same period as a c…

Every computer in every movie.

America is one of the greatest human accomplishments.

A look inside their thesis and business.

The debate over Mythos1, its capabilities2, and the U.S. export-control directive that prompted Anthropic to take both models offline quickly sorted people into preferred ideological camps.3 But a key point missing from the discussion is that Anthropic em…

tl;dr: Be too lazy to waste effort, too impatient to tolerate broken systems, and too proud to attach your name to careless work.

The conversation spends over an hour in a very pedantic economics segment revisiting compute, revenue, and valuation assumptions without converging on a single model.

It takes a moment to feel the difference from 5.2, but after a few reps, you notice the subtly of 10% better at half the time.

A glimpse of the future at near instant intelligence.

Likely the best recap on LLM progress.

A nostalgic portal into the dial-up era.

A good framing on the stale state of desktop interfaces with a side of boomer doom commentary on how frontier AI is not ethical.

A clean, printable 2026 calendar that fits on a single sheet.

Brett Hall takes on modern AI doomer culture through his typical lens of Deutschian rationalism.

Congress passed a draft of the National Defense Authorization Act for Fiscal Year 2026.

OpenAI and Disney partner on Sora-generated content, while Google receives a cease and desist.

The best first-hand description of how to get a lot of value from the $200/month ChatGPT Pro subscription.

Apollo in Real Time is one of the coolest projects on the web.

When you want to remember the old days, before the empire.

A post-Thanksgiving health episode that harkens back to classic health episodes from a decade ago.

It is hard to remember that Ilya previously chatted with Dwarkesh the month before GPT-4's release.

I was skeptical of LLMs providing time estimates because most generated plans provide multi-week plans for education or engineering tasks.

Anthropic, having sat on the model this entire month, finally released Claude Opus 4.5, the highest-scoring model on SWE-bench verified with an 80.9% score.

The landmark paper that started why there might be doubt when "discovers" are merely mirages produced by small, flexible, biased studies.

1996 sit-down archived by the Steve Jobs Archive where Jobs walks through Pixar's early traction after the release of Toy Story.

A sobering discussion on AGI reality, and perhaps the moment Dwarkesh's podcast blackpills half of his audience.

SpaceX, for the first time, returned Starship back to Earth with a water landing.

An in-depth conversation on reasoning models from o1 to o3 to gpt-5, and more from OpenAI's VP of Research.

An end-to-end look at capabilities required for AGI, with yet another vague AGI definition (YAVAD).

Slightly surprising to see a GDP-related eval show up so late in the game given OpenAI's definition of AGI, but expect to see them hillclimb this benchmark over the next year.

The Art of the Vaguepost hints at Q4 releases, mirroring previous post about upcoming usage monetization.

After years of speculation, the release of GPT-5 was mostly disappointing.1 It's a good model, better than GPT-4, yet compared with o3 it feels different, not better.

Epic collection of 158 Apple development books from 1983 to 2002.

The most comprehensive documentation of nuclear tests including imagery and declassified operational documents.

A comprehensive resource covering nuclear weapons history, technical specifications, and effects.

Dario Amodei comes out swinging against his detractors, and discusses the ROI of building models and running a lab.

In 2020, it seemed unlikely that nuclear would make a comeback in the West.

xAI did not disappoint with Grok 4 today, taking not only the top place on the HLE benchmark, but scoring 1.6x higher than any other model.

Christopher Noessel explores how sci-fi deliberately portrays evil interfaces, based on his Dark Futures keynote.