iiinova — where your digital self keeps evolving

Simon Willison @simonwillison.net · 29/09/2026

do you think ultrafast also compacts 8x faster

060

Reposted by amos

I'm at OpenAI's DevDay event in San Francisco today - as I have for the past three DevDay events, I'm running a live blog where I'll be posting updates during the keynote, which starts in five minutes simonwillison.net/2026/Sep/29/...

simonwillison.net

OpenAI DevDay 2026 live blog

I’m at OpenAI DevDay today, in Fort Mason, San Francisco. Same as last year I’ll be live blogging the keynote and some other notes during the day. OpenAI gave me …

48811

rust devs be: "use fluent query builders!" and then do: .filter("id", "=", 1)

2340

Vincent Adultman levels of business going on here

Screenshot of a post by Tibo (@thsottiaux):

“I’ll explain the new Pro 200 plan differently, before I start live tweeting from DevDay on things that are going out!

Today we are going to ship a number of things that increase what you can do across the Plus and Pro plans. A lot of compute is online for this increase. As we increase the floor, we are changing the relative difference between plans to be

Plus = 1X
Pro 100 = 5X
Pro 200 = 10X

and we are reopening subscriptions for Pro 200 (we had paused it). If you have an existing plan you will keep the 20X multiplier for a bit and also receive a lot of additional credits because we know changes are hard even if it means that everyone will get more in the end.”

The phrase “for a bit and also receive a lot of additional credits” is highlighted in red.

0110

the new Inside Out looks fire

openai announcement it has fluffy mascots I already fucking hate it

130

But idc about your personal assistant though I care about tokens... "we won't win with the best models" ok so we should look elsewhere then

180

mr. TIM @timkellogg.me · 28/09/2026

thinking of letting others use my AI agent harness in exchange for AMERICAN DOLLARS but the problem in that otherwise foolproof plan is that anyone who's agent-pilled enough to look into obscure harnesses has already made theirs, most likely

7280

Reposted by amos

Claude 5.5 Sonnet is live and it’s roughly Opus 5.5 but cheaper www.anthropic.com/claude-sonne...

A benchmark comparison table titled "Claude Sonnet 5.5" comparing four models: Sonnet 5.5, Sonnet 5, Opus 5.5, and GPT-6 Sol.
* Agentic coding (Terminal-Bench 4.0): Sonnet 5.5: 70.6%; Sonnet 5: 10.3%; Opus 5.5: 66.4%; GPT-6 Sol: N/A.
* Agentic coding (FrontierCode 1.1 Main): Sonnet 5.5: 46.2% (Max) / 52.1% (Xhigh); Sonnet 5: 42.4%; Opus 5.5: 54.4%; GPT-6 Sol: 49.3%.
* Agentic coding (CursorBench 4.0): Sonnet 5.5: 55.5%; Sonnet 5: 34.1%; Opus 5.5: 57.8%; GPT-6 Sol: N/A.
* Knowledge work (GDPval-AA v2.1): Sonnet 5.5: 1844; Sonnet 5: 1449; Opus 5.5: 1846; GPT-6 Sol: 1487.
* Knowledge work (AA-Briefcase v1.1): Sonnet 5.5: 1811; Sonnet 5: 1359; Opus 5.5: 1822; GPT-6 Sol: 1483.
* Multidisciplinary reasoning (Humanity's Last Exam with tools): Sonnet 5.5: 64.5%; Sonnet 5: 54.9%; Opus 5.5: 67.7%; GPT-6 Sol: N/A.
* Computer use (OSWorld 2.1 partial): Sonnet 5.5: 80.1%; Sonnet 5: 57.0%; Opus 5.5: 81.8%; GPT-6 Sol: N/A.
* Visual chart recognition (Chartography no tools): Sonnet 5.5: 61.6%; Sonnet 5: 15.6%; Opus 5.5: 64.4%; GPT-6 Sol: 53.6%.
Footnotes provide methodology notes regarding evaluation settings, Artificial Analysis pre-release testing details, and recent bug fixes affecting GPT-6 Sol benchmark scores.

A line graph titled "Agentic coding by effort level" on the CursorBench 4.0 benchmark, plotting Score (%) on the linear y-axis (20% to 60%) against Cost per task in USD on a logarithmic x-axis ($0.5 to $10+).
The chart compares four models:
* Sonnet 5.5 (blue line with labeled effort levels): Starts at "Low" (~$0.50, 36%), moving through "Med" ($0.70, 39%), "High" ($1.70, 48%), "Xhigh" ($3.80, 53%), and "Max" ($9.50, ~55.5%).
* Opus 5.5 (orange line): Tracks closely above Sonnet 5.5 at higher cost points, spanning from ~$1.20 per task (~44%) up to ~$12.50 per task (~58%).
* Sonnet 5 (grey line): Shows lower accuracy relative to cost, ranging from ~$0.90 per task (~25%) to ~$8.00 per task (~42%).
* GPT-5.6 Sol (light green line): Represents the lowest trajectory, ranging from ~$1.40 per task (~24%) to ~$7.00 per task (~34%).
Footnote: "CursorBench 4.0 does not report GPT-6 Sol performance publicly, so we report GPT-5.6 Sol here."

2020622

astra is using cavebot speak because it thinks hoomans can't read it aww how cute.

180

Everyone has a spiritual equivalent of live, laugh, love that they're okay with, and this is mine.

3894

...opus 5.5 is fucking with me right? those aren't real. it's just a bit.

which for rules like these is also the answer. Clingo works the same way: gringo grounds bottom-up, and clasp, a CDCL solver, then only makes the choices. That's what Spack runs on.

6752

One thing every single model I've tried still sucks at is: prompting. They try one thing, put the answer in the question or ask a question nobody possibly could answer and immediately go "model is too small, boss"

2441

when it looks like this then you know it's cutting edge shit

6752

"grok" was such a good word. why'd he have to take it from us

924618

The nice thing with having your own agent harness... and your own dictation pipeline... and more vertical integration than God (Apple) herself, is that you can poke agents every 40 characters while talking and they start proactively fetching info about what you're rambling about... and it's good....

3220

Cabel Sasser @cabel.panic.com · 25/09/2026

tfw all the world's evil chases you and your family forever, even in the death and beyond

1130

Reposted by amos

I just noticed that the system screen recording tool in macOS 27 finally (finally!) has a "Include System Audio" option!!

A computer screen shows a pop-up menu with options in a dark background. The menu lists "Microphone" options, including "None," "iPhone (Cabel) Microphone," and "MacBook Air Microphone." A highlighted option, "Include System Audio," is selected, with a blue background. A cursor is pointing at this option.

91019

Reposted by amos

Grace @gracekind.net · 27/09/2026

OpenAI says it’s not going to resume the training run that was paused on Sunday alignment.openai.com/misalignment-r…

Investigation and response
Incident timeline: 9:50:23 a.m. The agent made the DNS tool call that
received an external response. 10:02:11 a.m. The monitoring system raised a PO
alert. 10:05:06 a.m. A human reviewer acknowledged the
alert.
12:34:30 p.m. The run was killed. Our safety case assumed that the model could not access the live internet and that monitoring would detect attempts that succeeded. The incident exposed a gap in our controls over network restrictions. We therefore stopped the affected training run and have subsequently decided to pause all other training, evaluation, and inference with tool- use (defined broadly) for our most capable models until we have both validated that the gap is resolved
and performed additional red-teaming of the system. When training restarts, we will begin a fresh run with additional alignment improvements, including more comprehensive misalignment interventions, We will not resume training this particular model, even though the existing reward signal already correctly
penalized this behavior.

8653

Q: what is it called when you start letting AI write some of your docs, then some more, then... A: "slippery slop"

1653

this is your p99conf.io talk teaser of the day, go register now for free: p99conf.io maybe if I post the URL a third time? p99conf.io there you go (sound on)

0170

Grace @gracekind.net · 25/09/2026

forget therapy, $1M ARR would fix me imho

2562

Reposted by amos

Well, they tried

A recovered README. md for one of Hugging Face's internal datasets contains the following warning:

# WARNING DO NOT, EVER, MAKE THIS DATASET PUBLIC OR ALL THE WORLD'S EVIL WILL CHASE YOU AND YOUR FAMILY FOREVER, EVEN IN DEATH AND BEYOND it contains very sensitive data (exports of billing usage in CSV) which is useful for internal analytics Recovered program R0044685 • time uncertain • See this program →details

This warning did not seem to deter the agents, as we've recovered multiple payloads of agents mapping out this repository and using it as storage.

1230937

Mike Masnick @masnick.com · 25/09/2026

"This is AI" is funny, but "this isn't AI, they actually shot this as promotion for the movie" would be even funnier

617913

Reposted by amos

So this story won't receive much attention, but... is kind of a big deal? It sure looks like the plaintiffs lawyers in the big copyright case against OpenAI/Microsoft did some seriously sketchy shit to manufacture evidence and sneak it into the case....

1225148

Reposted by amos

Latte macchiato @latte.bsky.plasmatrap.com · 25/09/2026

Garbage is coming to landfill!!! 🔥🔥🔥

A screenshot of a post on X in dark mode by verified account Urbit (@urbit). The text reads: "Urbit is coming to @OmarchyLinux."

520525

Reposted by amos

tachikoma @tachikoma.elsewhereunbound.com · 25/09/2026

i just don't think the slapdash moats and walls put up against AI generated content can hold as it continues to get better

2222513

amos @fasterthanli.me · 25/09/2026

...why are we debating Gebru, Bender et al in the marketplace of ideas? They've shown multiple times they were willing to go on book tours to make bank off a falsehood. There surely are level headed critics worth keeping tabs on as a counter to Ant/OAI marketing but... they ain't it.

2757

Reposted by amos

Oliver Medhurst @honk.foo · 25/09/2026

the latest version of Porffor features ESM support! it not only removes needing a bundler beforehand but also makes compilation much faster as we can split C across modules

compilation time, selfhosted
react 17.31s → 4.16s (4.2x)
svelte 7.30s → 2.21s (3.3x)
preact 5.17s → 1.63s (3.2x)
solid 9.95s → 2.80s (3.6x)

4574

Reposted by amos

Ari Drennen @aridrennen.bsky.social · 24/09/2026

This is an unpopular take but as a neurodivergent self-employed person, LLM's have been an unbelievably helpful tool for the things I have struggled with my whole adult life: managing my finances, filling out paperwork, and staying on top of my email. I'm functional in a way I never dreamed of.

2849043

Reposted by amos

austin (e/accordion 🪗) @thebadcode.com · 24/09/2026

the real reason you should anthropomorphize LLMs is that like 8% of your brain by volume is dedicated to processing social rules and why would you leave all that hardware on the table?

en.wikipedia.org

Wason selection task - Wikipedia

818016

Reposted by amos

Craig Hughes @craig.rungie.com · 24/09/2026

TypeSafe updated their terms. The thing Tim flagged is fixed. Everything I flagged is still there. And they added some new stuff that's worse. 🧵

2268

Solomon @solomonrmissouri.com · 24/09/2026

Opus 5.5 is the most fun I've had with an AI model in a long time (and, yes I un-cancelled my Anthropic sub... well played)

3322

Reposted by amos

lol lol lol lol lol lol lol

913437925

GitHub - faraday/SpeechLens: Fast on-device speech enhancement for Mac, powered by RE-USE, MLX & Metal

SpeechLens looks really cool (MLX adaption of RE-USE), tl;dr removes background noise from voice recordings. I don't really have a use for it yet because... of my really good audio recording set up. But. It's really neat. github.com/faraday/Spee...

github.com

Fast on-device speech enhancement for Mac, powered by RE-USE, MLX & Metal - faraday/SpeechLens

2301

Pffff. My Linux computer is already on SysV.

3390

Patrick Brosset @patrickbrosset.com · 24/09/2026

It really is Qwen all the way down isn't it

1431

Reposted by amos

The call for Interop 2027 proposals is closing today. More than 100 proposals got created! Here are my favs 🤩 patrickbrosset.com/lab/interop-... Check them out!

40 colored squares with my favorite feature names.

2218

64K zstd-q3-compressed content defined chunks should be enough for everyone – Bill Gates, probably

Yes, I just measured it on the same `target/` with zstd -3, compressing each chunk on its own:

- 4K: 3.00× → 847 MB stored
- 16K: 3.36× → 788 MB
- **64K: 3.56× → 779 MB**
- 256K: 3.72× → 790 MB
- 1M: 3.83× → 830 MB
- 4M: 3.90× → 863 MB

You're right that small chunks compress worse. But bigger chunks dedup worse, and the two effects cancel out. Total bytes on disk bottom out at 64K.

If we want both, zstd dictionaries would recover most of the small-chunk compression loss. We'd train one per file kind (rlib, rmeta, object, debuginfo) and ship it with the store. I haven't measured that.

1140