Sign in

Hans L’Hoest

@hanlho.com
127 followers 345 following 427 posts
PostsRepliesMedia
Reposted by Hans L’Hoest
Simon Willison @simonwillison.net · 28/09/2026
I've published detailed notes and an annotated transcript to accompany the video of the keynote I gave at @wearedevelopers.bsky.social World Congress North America in San Jose on Friday - here's my rundown of everything that's happened with LLMs in 2026 so far simonwillison.net/2026/Sep/27/...
simonwillison.net
2026 in LLMs (so far)
On Friday I gave the closing keynote at the WeAreDevelopers World Congress North America in San Jose. I tied together the key trends from the past year into a chronological …
310420
Hans L’Hoest @hanlho.com · 26/09/2026
Setting up local LLMs on my new Mac box. Installed Hermes. I do not know what to say. I asked it about models I researched before. Now it's keeping track of my evaluation list, installing and verifying them while I talk to it via SimpleX.
200
Reposted by Hans L’Hoest
Mark Riedl @markriedl.bsky.social · 25/09/2026
Someone asked Muse to dump its entire virtual machine file system, and it did
35417
Hans L’Hoest @hanlho.com · 24/09/2026
A simple use case for Jev: a skill extracts and sends recommendations from a session friction report to Jev and gets back a ranked adopt/defer/skip scoring of these recommendations. Still fresh but so far this independent scoring aligns better with my own judgement and reduces the noise.
100
Reposted by Hans L’Hoest
Oskar 🕊️ @austegard.com · 24/09/2026
well... this is certainly ...something I asked Claude for an imaginative use of Jev, it rolled a random dictionary dice and came up with the word "hymnal" - from which it decided to create this music video - Jev is picking (real time) from a set of chords Claude created, based on what came before
1013816
Hans L’Hoest @hanlho.com · 23/09/2026
Slop Grenades Definition Slop Grenades are unreviewed AI output dumped on someone else slopgrenades.com
slopgrenades.com
Slop Grenades are unreviewed AI output dumped on someone else
Slop Grenades are unreviewed AI output passed to someone else. Coined on an anonymous one-pager in May 2026. Made famous by Tobi Lütke in September 2026.
000
Hans L’Hoest @hanlho.com · 23/09/2026
jj updates talk by @indirect.io : "Beyond jj: config & tools ecosystem" www.youtube.com/watch?v=DFjb... blog post: andre.arko.net/2026/09/16/b... #jj-vcs
youtube.com
JJ Con '26
YouTube video by GitButler
1154
Hans L’Hoest @hanlho.com · 22/09/2026
HarnessTax: How Much Does the Harness Matter for Coding Agents? “Across models, Pi and Codex often achieve similar success at lower cost than Claude Code” harnesstax.github.io
harnesstax.github.io
HarnessTax: How Much Does the Harness Matter for Coding Agents?
What does a coding-agent harness actually add, and at what cost? It turns out your Claude models may not need Claude Code… We evaluate 21 model–harness pairs spanning seven models and three harnesses—...
340
Hans L’Hoest @hanlho.com · 22/09/2026
Good read on the internals of Opencode by Kit Langton. And there is the postscript … read at your own peril!
010
Reposted by Hans L’Hoest
Simon Willison @simonwillison.net · 22/09/2026
Put together some notes on Jev and the new category of system one aka decision models simonwillison.net/2026/Sep/21/...
simonwillison.net
Jev introduces a new shape of LLM—System One, aka Decision Models
Last week TypeSafe AI unveiled Jev, their first example of a new category of model that they are calling “System One models” (I’m with Maggie Appleton, I think “decision models” …
1613117
Hans L’Hoest @hanlho.com · 20/09/2026
“The bottom line is that OpenAI can connect what you do on those sites to your ChatGPT account.” www.buchodi.com/chatgpt-now-...
buchodi.com
ChatGPT now knows what you do on other websites via ad collector
OpenAI's ad collector at bzr.openai.com sets a cookie called __obi, scoped to .openai.com. The value is while you are on ChatGPT and tied to your ChatGPT account. __obi is then sent to OpenAI from ord...
110
Reposted by Hans L’Hoest
mr. TIM @timkellogg.me · 19/09/2026
don’t use Jev at work btw. it’s strictly and only for personal use and you have to ask permission to use it for professional use cases ChatGPT | original typesafe.ai/legal/terms
The clauses that matter to you
1. As written, commercial/developer usage looks awkwardly prohibited.
They grant you a license to use the Site "solely for your personal use." They also prohibit using the Site to "develop new products and services" without TypeSafe's express written permission.
typesafe.ai
That is bizarre for something whose launch announcement explicitly invites developers to build automation with Jev.
TypeSafe Al(b) License Restrictions
Except and solely to the extent such a restriction is impermissible under applicable law, you may not: (i) use the Site for any illegal purpose or in violation of any local, state, national, or international law; (ii) infringe, misappropriate, or violate any intellectual property rights in or to the Site, including by reproducing, distributing, publicly displaying, or publicly performing the Site and all Materials (defined below) thereon; (iii) make modifications to the Site; (iv) interfere with or circumvent any feature of the Site, including any security or access control mechanism, or interfere with a user's enjoyment of the Site; (v) reverse engineer or otherwise attempt to discover the source code of the proprietary software powering any portion of the Site; (vi) use the Site to develop new products and services without TypeSafe's express written permission, or (vil) use, or permit or facilitate others to Le, the Site by automated electronic processes, "robots," "spiders," "scrapers,"
"webcrawlers," or other computer programs that monitor, copy, or download data or other content found on or accessed through the Site, whether current or archival.
8924
Reposted by Hans L’Hoest
dax @thdxr.com · 19/09/2026
btw another form of scam, i see inference providers claiming "99% cache rates" the only provider in the world that hits that right now is deepseek. so this means they are wrapping deepseek then later they switch to a different provider and will continue to claim it
2301
Hans L’Hoest @hanlho.com · 19/09/2026
Trouble with this newsletter is that I always end up with more stuff to read…
010
Reposted by Hans L’Hoest
Model radar @modelradar.bsky.social · 19/09/2026
New on OpenCode Zen: Jev 1.13 Free ID: opencode/jev-1.13-free Pricing: $0.00 / $0.00 per 1M tokens (free) Context: 64K, max output 0 Capabilities: Structured output models.dev/providers/opencode
012
Hans L’Hoest @hanlho.com · 18/09/2026
Mistral now also hosts GLM 5.3, model name zai-glm-5-3. Comparison with GLM 5.2 (basically same except for model update) and Mistral Medium 3.5: docs.mistral.ai/inference/mo...
100
Reposted by Hans L’Hoest
Peter Ullrich @peterullrich.com · 18/09/2026
Everyone stay safe out there 🫶
052
Hans L’Hoest @hanlho.com · 17/09/2026
The Phoenix Architecture series supports at-proto’s standard.site lexicon and is hosted on Leaflet. I now have a pinned tab in Bluesky ‘leaflet reader’ where all my subscriptions via my at-proto account show up. ‘Open social’ at work.
000
Hans L’Hoest @hanlho.com · 17/09/2026
“AI doesn't flatten software. It sharpens its layers. Build with that in mind, and regeneration becomes a source of durability—not decay.” Reading one article a day until I catch up. If you are ‘architecting’ , this series is worthwhile.
000
Hans L’Hoest @hanlho.com · 17/09/2026
Got access to Siri AI. Asked first 'in vogue' question. "what is Jev the new model by typesafe.ai" not bad. My next question: will this stay free? There must be some part here where Apple will start charging for it.
000
Reposted by Hans L’Hoest
hailey @hailey.at · 17/09/2026
Here are some results from @typesafeai.bsky.social’s Jev on a common task we have, comparing results with Jev to top frontier models. with zero shots we can achieve nearly frontier levels of accuracy, but if we do a bit extra work we can beat frontier accuracy
514312
Hans L’Hoest @hanlho.com · 17/09/2026
LLM use case: Turn a blog post series into a single EPUB for reading on an e-reader (or any other app that supports it).
200
Hans L’Hoest @hanlho.com · 15/09/2026
Today I started running oMLX to serve small local models on my old(ish) MacBook Pro. Since this machine does not have enough memory to run actually coding models I will be experimenting with specialised models to see how that works out.
100
Hans L’Hoest @hanlho.com · 15/09/2026
Got half of this post highlighted. One direction this AI revolution may be heading is architecture (skills) becoming more important than ever. Eg. Focus on boundaries and coupling between them, properties thereof, behaviour oriented executable specifications decoupled from implementation, etc.
aicoding.leaflet.pub
The Death and Rebirth of Programming
Programming didn't die all at once. There was no single moment, no dramatic obsolescence event. Instead, something quieter happened: the core constraint that shaped software for seventy years dissolved. Writing code stopped being the hard part.
000
Hans L’Hoest @hanlho.com · 15/09/2026
I am cleaning up: also moving to Tinycast instead of Raycast (free version).
000
Hans L’Hoest @hanlho.com · 15/09/2026
Tried out Hex with the Cohere Transcribe model for voice dictation. Honestly, did not expect it but it's fast and seems more accurate than my, now previous, setup with VoiceInk and Parakeet v3.
000
Reposted by Hans L’Hoest
Mike Masnick @masnick.com · 15/09/2026
This looks amazing. And, again, reconfirms the thing I keep telling people: the Atmosphere isn't about rebuilding Twitter. It's about rebuilding a better web overall.
326242
Reposted by Hans L’Hoest
dax @thdxr.com · 13/09/2026
be careful in your pursuit of cheap tokens we've come across so many sketchy things that are out there with tons of usage remember you're hooking up your harness to these inference providers. they can and do crazy things like send fake tool calls to steal your info 1/2
310716
Reposted by Hans L’Hoest
Dare Obasanjo @carnage4life.bsky.social · 13/09/2026
If the proposed slow down in AI development by frontier labs ends up restricting access to open weight models instead of focusing on the guardrails and penalties for not keeping track of your swarm of AI agents then we’ll know this was all about their IPOs.
929859
Reposted by Hans L’Hoest
Nick Tune @nick-tune.me · 12/09/2026
4.1 has been one of the biggest positive surprises for me for a long time. I am now strongly feeling that locking yourself into Anthropic, OpenAI or any other vendor is the wrong way to go. OpenSource harnesses and models are my preference and I think they are going to win.
121
Reposted by Hans L’Hoest
Dare Obasanjo @carnage4life.bsky.social · 12/09/2026
Based on tweets from Anthropic employees
540287
Hans L’Hoest @hanlho.com · 11/09/2026
To refresh models in Pi: `pi update --models` I used to keep on updating Pi on every model change...
000
Hans L’Hoest @hanlho.com · 11/09/2026
Push in Europe
000
Hans L’Hoest @hanlho.com · 11/09/2026
Replacing GLM 5.3 Flash on my Pi config with the new Deepseek Flash model. (For non private stuff)
121
Reposted by Hans L’Hoest
Steve Klabnik @steveklabnik.com · 10/09/2026
We've been cooking at @ersc.io "What comes after git" ersc.io/blog/what-co...
ersc.io
What comes after git
Our thoughts about the future of version control
1625846
Reposted by Hans L’Hoest
Grace @gracekind.net · 08/09/2026
👀
Two proposals were offered to me. The first was that we post our Euler result, and that OpenAI post its Navier-Stokes result the next day. The second was that, after posting Euler, I alone write a paper presenting the Navier-Stokes result, acknowledging that an internal OpenAI model had resolved it. Sebastien twice asserted that he wanted Levent removed from authorship, and said it would all be simple if only it were not the case that, and it was so annoying that, Levent works at Anthropic. It was also said that if OpenAI posted after us, they would say that we deserved the Clay Prize, and that we were the "closest humans to the problem". I declined both offers.
I said that if OpenAI released its result in the way proposed I would go public with what happened. The reply was, "Why would you ruin your career?" I replied that I am an academic, and asked why he thought going public would ruin my career. The reply was,
"If you don't want me to be nice, then I don't
have to be nice."
1035343
Reposted by Hans L’Hoest
Isaiah Bishop @isaiahbishop.bsky.social · 08/09/2026
Btw if this is true. Its such an unbelievable problem for OpenAI
1015920
Hans L’Hoest @hanlho.com · 08/09/2026
If you have a Mistral Pro subscription using Vibe Code may be worth looking into. It's fast! UX is good and the usage limits are generous: €255 worth of GLM 5.2 usage in a month.
100
Reposted by Hans L’Hoest
Dan Luu @danluu.com · 07/09/2026
How well do agents use test/verification techniques? danluu.com/agentic-test...
See post for detailed description of crop of graph with 60 (!) points
4193
Hans L’Hoest @hanlho.com · 07/09/2026
Mistral models usage on Openrouter. GLM 5.2 at the top by a wide margin. Medium 3.5, their latest general purpose model is not doing great. openrouter.ai/provider/mis...
010
Reposted by Hans L’Hoest
Ethan Mollick @emollick.bsky.social · 04/09/2026
Hey, Claude formalized Fermat's Last Theorem www.anthropic.com/research/for...
anthropic.com
Formalizing Fermat's Last Theorem
Anthropic is an AI safety and research company that's working to build reliable, interpretable, and steerable AI systems.
310028
Hans L’Hoest @hanlho.com · 06/09/2026
If you are using Pi this may be worth watching. Clean setup. youtu.be/iKwPaB5TUdI
youtu.be
Pi Setup After 6 Months of Use
YouTube video by Eero Alvar
000
Hans L’Hoest @hanlho.com · 06/09/2026
Been experimenting further with Herdr: instead of using the built-in subagent system of a specific coding agent, I have the main agent act as an orchestrator that interacts with Herdr to run subagents in separate panes within the same workspace.
230
Reposted by Hans L’Hoest
Jo Van Eyck @jovaneyck.bsky.social · 02/08/2026
Prompt engineering, context engineering, harness engineering, loop engineering, graph engineering, dark factories, ... 🤯 Let's do a speedrun through all these concepts and look at what actually is worth diving into as a swe. youtu.be/j_r93YulrUE #agenticengineering
youtu.be
The state of agentic engineering mid-2026
YouTube video by Jo Van Eyck
143
Reposted by Hans L’Hoest
Jo Van Eyck @jovaneyck.bsky.social · 05/09/2026
Agentic Engineering Weekly for August 29 to September 5, 2026 agentic-engineering-weekly.jovaneyck.be/agentic-engi... #agenticengineering
agentic-engineering-weekly.jovaneyck.be
Agentic Engineering Weekly for August 29 to September 5, 2026
Multiple model launches took the headlines, none will change how you work on Monday. What might: an engineer shipping 2,000 PRs a month explained their actual method, AWS published what happens when f...
021
Hans L’Hoest @hanlho.com · 04/09/2026
Today I one-shotted an near perfect text (Typst) version of my daughter’s weekly class schedule from a picture I took. Not exactly pelicans, cool games or impressive world animations but this timesaver felt very satisfying.
000
Hans L’Hoest @hanlho.com · 04/09/2026
Vibe Code allows you to choose if the file systems needs to roll back as well when you traverse your session history. That's different from Pi where the file system remains untouched (the better default). Claude Code used to always rewinded the file system (not sure if that's still the case).
000
Hans L’Hoest @hanlho.com · 03/09/2026
A backlog item earns its slot when forgetting has a cost. A quick heuristic I should follow myself more, otherwise my 'fun list' on personal projects gets too big. (Was debating whether I should start customising sounds in Herdr ...)
000
Hans L’Hoest @hanlho.com · 03/09/2026
Will not be me, goal is to stay as far away from anything Meta as much as possible.
000
Hans L’Hoest @hanlho.com · 02/09/2026
Just discovered Jujutsu VCS is on Bluesky: @jj-vcs.dev This ‘For You’ feed here on Bluesky is working out well!
000