←
Answer.AI64 posts
Answer.AI avatar

Answer.AI

@answerdotai

A new kind of AI R&D lab which creates practical end-user products based on foundational research breakthroughs

◷ Joined March 2024▧ 1 Photos and videos
Jeremy Howard avatar
⟳ Answer.AI retweeted
Jeremy Howard@jeremyphoward·Nov 9

This is one of the most important things for startup founders to know about. If you're using your investors' preferred lawyer, then they're working for your investor, not for you. Use tryvirgil.com/ instead.

VC: this deal would go much more smoothly if you used our preferred lawyer First time founder: yeah of course, let’s not rock the boat here. Grateful to have you aboard VC: this deal would go much more smoothly if you - 2nd time founder: *unholstering gun*

David Gwyer - AI & ML Researcher/Engineer avatar
⟳ Answer.AI retweeted
David Gwyer - AI & ML Researcher/Engineer@dgwyer·Oct 1

Control Excalidraw Instantaneously in SolveIt via Jev! Here's an example of just how fast Jev is in making decisions. Updating an Excalidraw scene in SolveIt via Jev is almost instantaneous! Each separate Excalidraw prompt type English command is sent to Jev for evaluation and then an action is chosen and executed.

2:19
Pol Avec avatar
⟳ Answer.AI retweeted
Pol Avec@pol_avec·Aug 31

New tiny package: github.com/AnswerDotAI/shot2… Press a hotkey, drag over any region of your screen, and whatever's written there is on your clipboard. Including text you can't select: screenshares, videos, images. Setup is `pip install shot2text` + 2 config lines. The whole library is under 50 lines of Python. That's possible because of two new @answerdotai libs, macmage and imp, which open up macOS scripting from Python. What they do, in 🧵

Chris Lattner avatar
⟳ Answer.AI retweeted
Chris Lattner@clattner_llvm·Aug 11

Mojo 1.0 is here! This is a huge milestone for the community at mojolang.org ! Mojo reaches language stability and its first epoch of feature maturity. Tune in to ModCon next week to learn what this unlocks! 🚀

Mojo 🔥 1.0 is here. Since we open-sourced the standard library, nearly 200 contributors have landed more than 1+ pull requests, changing 200k+ lines of code, and more than 1k+ others have filed issues that shaped the language. To every developer who filed an issue, opened a pull request, wrote a language proposal, or built a package: thank you for being the architects of this language alongside us. Full blog post: modular.com/blog/modular-26-…

Scott H. Hawley avatar
⟳ Answer.AI retweeted
Scott H. Hawley@drscotthawley·Aug 6

Excited to share a new Preprint & LIVE DEMO! This is >2 years of my life building fast, lightweight CPU-friendly models to facilitate songwriters' iterative workflows, with Representations usable for Understanding and Generation, trained in a Self-Supervised way. First, the demo:

Post attachment
David Gwyer - AI & ML Researcher/Engineer avatar
⟳ Answer.AI retweeted
David Gwyer - AI & ML Researcher/Engineer@dgwyer·Aug 1

(1/8) Why do eigenvalues and eigenvectors appear so often in machine learning? Because they uncover important directions hidden inside data and transformations. Let’s follow that idea from geometry to PCA, clustering and optimization.

Benjamin Warner avatar
⟳ Answer.AI retweeted
Benjamin Warner@benjamin_warner·Jul 30

Congratulations to Liquid on a strong open-source encoder release. Looks like ModernBERT has some competition😄 My read from Liquid's benchmarks: - for multilingual try mmBERT, ModernBERT, & LFM-E - for english-only try ModernBERT & LFM-E

Today we release LFM2.5-Encoder-230M and LFM2.5-Encoder-350M: bidirectional encoders that stay fast at long context, even on CPU. > LFM2.5-Encoder-230M: about 3.7x faster than ModernBERT-base on CPU at 8,192 tokens. Under 30s per forward pass, versus over a minute and a half. > LFM2.5-Encoder-350M: 4th of 14 models on GLUE, SuperGLUE, and multilingual classification, behind only three larger models, one of them nearly 10x its size. 🧵

2:16
David Gwyer - AI & ML Researcher/Engineer avatar
⟳ Answer.AI retweeted
David Gwyer - AI & ML Researcher/Engineer@dgwyer·Jul 25

A tiny equation (Clifford attractor) can make a surprisingly complex plot. Start with a point, apply the same simple rule again and again, and after enough repetitions the points begin to trace out this delicate structure. Not randomness. Just iteration.

Post attachment
Pol Avec avatar
⟳ Answer.AI retweeted
Pol Avec@pol_avec·Jul 16

Very interesting paper. It is framed as: model intelligence is inversely proportional to # of human interventions However, I think that having a human might give a higher quality result, even if you could do it with less interventions. I used Solveit to implement & score myself on one of the tasks (very easy & reproducible paper setup, kudos for that! It was a relatively simple data anonymization task, my score was perfect (same as the AI). The resulting code was much simpler, almost half the LoC and methods. The median score of GPT 5.5 is 0.763. I would love to see more people's results! If you are interested in trying out a task, you can just use my dialog which is fully setup and just choose a different task! share.solveit.pub/d/62a08cb8… Share the results if you do

Post attachment

Introducing SWE-Together: a multi-turn benchmark built from real user–agent coding sessions. Coding agents are often benchmarked like exam-takers: given the full spec up front, then graded on the final code. But real coding help is a conversation — users clarify goals, add constraints, and correct course along the way. SWE-Together turns real coding work into a reproducible, verifiable benchmark: 109 repo-level tasks curated from 11,260 recorded sessions, replayed with a reactive LLM user simulator that preserves the original user’s intent. We evaluate agents as collaborators, not just patch generators: final pass rate and how many user interventions were needed to get there. In this evaluation snapshot, claude-opus-4.8 currently leads among the 7 agents we tested — achieving the highest pass rate while requiring the fewest user interventions. 📄 Paper: arxiv.org/abs/2606.29957 💻 Code: github.com/Togetherbench/SWE… 🌐 Website: togetherbench.com

Post attachment
Pol Avec avatar
⟳ Answer.AI retweeted
Pol Avec@pol_avec·Jun 17

I'm happy to say that today I'm joining @answerdotai ! I am happy at a technical, personal and spiritual level. On a technical level, they write fantastic software (SolveIT, FastHTML, nbdev, fastspec), articles (why you shoud write) and great courses (fastai, solveit). I see no end to my learning and path to mastery in here. On a personal level, I share interests with many of them beyond sw like sci-fi, chinese or robotics. On a spiritual level, AAI's stated goal is human flourishing, but in all fairness it seems to be more or less the goal also for all companies striving for AGI. So why AAI? When I read Huxley's "End & Means" I realized that the ends cannot justify the means. Moreover, using means opposed to your goal are unlikely to bring that goal to happen. Violent revolution for example, has rarely brought subsequent peace the book argues. Many labs have great goals too like human flourishing, long-term benefit of humanity, or improve billions of lives. However, often the means seem to sue each other, refusing to hold hands, or benefit humanity (but not China). Or lock in people to their sw, sometimes leveraging dark-flow patterns or fully gatekeeping access. I assume it's very hard to run such behemots in a fully consistent way. That's not a problem for AAI. It is a small team and their founders fully embody the company ethos. @ericries taught a whole generation of founders how to build w/ the Lean Startup. However he also saw that not all means are created equal: not all ways to make money are equally good. To make sure money making would be aligned with humanity's benefit he went out of his way with projects like the long term value stock exchange. More recently, his new book Incorruptible is literally a manual to make sure that companies don't go from creating value to society to extracting value from it. @jeremyphoward constantly works to enhance human creativity. He has pushed for making AI accessible to everyone through fastai courses and making sure no ludicrous amounts of hardware are needed to follow them or do meaningful research. The SolveIT discord is also one of the nicest places on the internet. The only rule is to try to be as helpful as possible, and he is the best example of that. I once heard Alan Watts say that there's no point in planning the future, if you can't enjoy the present. It seems to me the same is true for ends and means. Too many people seem to be living in borrowed time until AGI arrives disregarding their current learning, skills and also enjoyment. If you are one of those and you are not happy with that state of affairs, you are not alone! reach out

Salman // 萨尔曼 avatar
⟳ Answer.AI retweeted
Salman // 萨尔曼@ForBo7_·Apr 18

Going thru the lesson 9a notebook (by @johnowhitaker) of the fastai course, and created a little tool to play around and visualize the latents produced by an autoencoder Crafted with SolveIt and FastHTML (both by @answerdotai)

5:22
Jeremy Howard avatar
⟳ Answer.AI retweeted
Jeremy Howard@jeremyphoward·Apr 10

This is a great discussion. We've spent 2 years building a solution that's working well for us -- co-writing software side by side the AI in an notebook-ish environment. We call it the "solveit method". (We've created a course and platform for it: solve.it.com/ )

My colleague @istoica05 and I have been debating the role of specification in AI. I have argued that a key advantage of AI is that we can leave large parts of the specification unwritten. @istoica05 argued we need to focus on more specification. We converged on iterative disambiguation as the driving mechanism for specification. @vsreekanti and I have a new blog exploring this challenge more deeply and why maybe that interative process is still so important and still the bottleneck.

Pol Avec avatar
⟳ Answer.AI retweeted
Pol Avec@pol_avec·Apr 5

I am sharing this pi-notebooks extension for those using @badlogicgames Pi w/ notebooks. Notebooks have a lot of json-boilerplate and they are finicky to edit w/ regular text tools. However, their actual structure makes them easier for AI to edit if given the proper tools (ie they have cells ids, you can get a cells' summary, overview etc...) I have just ported the SolveIT / notebook editing tools from @answerdotai dialoghelper repo. It keeps context shorter and edit much better. GH repo: github.com/kafkasl/pi-notebo…