Daniel Gafni retweeted
I've discovered something more important than the Riemann Hypothesis: if you want your type checker to be faster on Pydantic, just tell it you are MYPY
Daniel Gafni retweeted
i caught up with @CompleteSkeptic last week to cover a few important things:
- the pink coat and hair and everything
- the AI bubble
- SF's crisis of purpose
00:00 - What’s up with the pink?
01:48 - Posting on X dot com
04:51 - AI has been a bubble for a while now
07:55 - Why are labs IPOing now?
10:21 - A lab, a neolab or a third thing
13:28 - The lab revenue problem
17:42 - SF’s crisis of purpose
25:37 - Not worrying about the macroeconomic climate
30:24 - The labs make a lot of bad bets
34:59 - Innovating on financial engineering
37:05 - Making employees arms bigger
38:23 - Pushing the frontier of craziness and bravery
Nice DeepSWE score! Does it survive a lowercase filename?
Introducing MaxxBench: 89 DeepSWE tasks with small changes to the instructions or starting code. Same engineering problem, slightly different requirement.
Models clustered around 74% on DeepSWE end up 23 pp apart.
Good at coding, or good at DeepSWE? 🧵
Daniel Gafni retweeted
Why don't we at TypeSafe call Jev a decision model?
If you think of Jev's questions as making *decisions*, I believe you'll use it at the wrong layer of abstraction and your reliability will suffer.
I wrote all about this on our site 🧵👇
Daniel Gafni retweeted
Why hasn't AI automated more work? What's holding coding agents back? And are the labs' valuations built on the right expectations?
@CompleteSkeptic, co-founder and CEO of @typesafeai, was the perfect person to discuss this and more. Diogo coauthored OpenAI's InstructGPT paper, and a few weeks before we recorded, TypeSafe launched Jev, a "system one" model that returns structured decisions instead of text.
It was a ton of fun to discuss this with Diogo.
The conversation includes:
0:00 Intro
1:14 The Early Bet on RLCD
4:40 Built for Humans vs Computers
6:39 Roadmap & Model Limitations
11:04 Mental Models of Models
16:07 Do LLMs Have a Place in 10 Years?
17:32 New Model Shapes
21:24 Problems Still Unsolved
27:00 The Tyranny of the KV Cache
31:21 Rethinking Coding Agents
34:40 Multimodality
37:52 Images, Games & Computer Use
39:36 Could This Have Been Built 5 Years Ago?
42:01 Benchmarks & Copycat Releases
44:12 Fundraising
45:27 Quickfire
Listen here:
YouTube: youtu.be/ham8dROHCrc
Spotify: bit.ly/4kyeNBQ
Apple: bit.ly/3WppmgK
Daniel Gafni retweeted
Announcing Dagger for Entire.
Goodbye pull requests! @GoEntire is a new Git forge for agents and humans. Now with native Dagger support.
Agents run the same checks before and after push, with a shared cache and without melting your machine.
dagger.io/blog/dagger-for-en…
Daniel Gafni retweeted
One thing that can be tricky to figure out is how to use Jev's primitives to handle more than one intent in a single user query.
My favorite way to approach this is what I call the "shopping cart" pattern: recursively pick the next thing until Jev says we're done.
I've personally used this to make a coffee ordering bot for our company Slack (using @DoorDash's CLI!), as well as a fast multi-intent handler for my smart home system.
Give it a shot!
Daniel Gafni retweeted
Announcing Dagger for GitHub.
Run the same checks before and after push. See results in your terminal and PR.
Cut out the middleman: just GitHub and Dagger Cloud.
Stop wasting compute: checks are cached, and you pay for exact usage, not machines.
dagger.io/blog/dagger-for-gi…
Daniel Gafni retweeted
One of my favorite Jev patterns is "line-by-line search", where you put an entire document in state, and ask a question for each line (or paragraph). This is super useful when you want to find the most relevant line for a query.
Use nouls when you expect one dominate answer, use scores when you want to highlight lines by overall relevance. The former is good for answering a specific question while the latter is good for finding all relevant passages.
Daniel Gafni retweeted
NEW: Top SaaS Vendors on Ramp (Oct 2026)
We clearly see AI spend pulled in two different directions.
On the downside, new and cheaper models are pushing AI spend down. Jev captured an entire percentage point of market adoption within a month (unheard of growth). And routers continue to gain.
On the upside, the increased adoption of expensive image and video generation models will push AI spend up, like Higgsfield and Runway.
Daniel Gafni retweeted
despite the unfortunate timing, we take disclosures seriously at @typesafeai
at approximately 6:49PM this Friday evening, we became aware of a massive unauthorized leak of ALL customer noul outputs
in effort to maximize transparency, here is the link: gist.github.com/diogo149/810…
Dagger is one of the sickest technologies I've seen in the recent years. It powers all our CI at TypeSafe! Can't recommend it enough
Announcing Dagger 1.0 beta.
Replace Jenkins and GitHub Actions with CI that matches the speed and scale of coding agents without wasting compute or tokens.
Featuring: test-native CI, dev shells, local and remote execution, official modules, and more.
dagger.io/blog/dagger-10-bet…
Others can benefit from some our modules that are open source!
Featuring the uv module specifically, it can build most Python packages out of the box. It has efficient caching, including local deps in monorepos.
github.com/typesafe-ai/dagge…
Daniel Gafni retweeted
Announcing Dagger 1.0 beta.
Replace Jenkins and GitHub Actions with CI that matches the speed and scale of coding agents without wasting compute or tokens.
Featuring: test-native CI, dev shells, local and remote execution, official modules, and more.
dagger.io/blog/dagger-10-bet…
Our latest real-time avatar model is live 🚀
Cara-4.1 Express builds on top of cara-4 to make it faster and more expressive. Tested blind against Tavus, HeyGen and LemonSlice, with the code public.
Try it: go.anam.ai/MrjU4Iz
xitter.cf/benatanam/status/21125…
Cara-4.1 Express is now live at @Anam__AI
Faster and more expressive than cara-4 with better lip sync and stronger prompt adherence. It also handles much wider framings for custom avatars.
We ran blind human evals: 85% of raters scored Anam higher than HeyGen, 88% higher than Tavus and 92% higher than LemonSlice. The eval code is public and reproducible.
Try it in the Anam Lab: go.anam.ai/dXLpQWy
Daniel Gafni retweeted
On an internal test set we generated for this task, Jev 1.13 had the highest accuracy and lowest cost when compared to other Decisions models and Claude Code's usage of Sonnet 5.
Daniel Gafni retweeted
New @typesafeai Jev capabilities!
Some decision models can ask questions about images but can they ask questions with images?
🧵 How to get the most out of the new Jev launch
Daniel Gafni retweeted
Happy JevDay!
But before announcements we have to talk philosophy…
The needs of early adopters / tinkerers / dreamers are quite different from what is needed for extremely stable infrastructure.
For the Jev release, we erred on the latter, but this release is for the former - with a gate so people can opt-in to the danger zone!
Daniel Gafni retweeted
today sana is becoming the new home for workday.
workday actions now take a few seconds to complete with dynamic ui.
powered by jev.
it's been v fun working with the @typesafeai team on this!
Daniel Gafni retweeted
launched a model
accidentally served trillions of tokens a day
29.4% of the fortune 500 showed up
raised a really big series A from @a16z
it's been 3 weeks
we would like to sleep now


























