tried going as much in detail as possible for the Colpali blog, it will be a long read antaripasaha.notion.site/Col…

Whether you are GPU poor or GPU rich, today's release of PyLate has something for you! GPU maxxers: MaxSim kernels greatly speed up training while lowering the memory requirements CPU enjoyers: TACHIOM enables lightning fast multi-vector indexing and search directly on CPU

Very excited to release late-interaction-kernels (LIK): fused Triton kernels for MaxSim, the scoring step behind ColBERT, ColPali & LateOn. 🚀 Numerically equivalent to PyTorch at a fraction of the memory, with day-0 support in PyLate & colpali-engine. (1/N 🧵)

the kind of exciting research work perplexity has been doing is amazing to see will be diving in for understanding this better

super cool work by @tomaarsen very excited to see how it’s extending the work done by lighton with modernbert and mixedbread with their latest release of pointwise rerankers the results already look promising with just 17M model open-source is the way to go🤝

a very happening reset with lego, books and a trip

The new generation of open state-of-the-art single and multi-vector retrieval models is here It's time, DenseOn with the LateOn 🎶 @LightOnIO releases models that leap past existing ones, and everything you need to do the same!

not sure what type of marketing was this but all the people who went crazy after the blogpost, quote tweeted and stuff and behaved like this is next level breakthrough. now you know they don’t understand shit about memory and are sloppy ai writers who just quote tweets on anything and everything

collection for next 3 months

highly agreed droid is by far the best agent out there. having been using for the past 5-6 months and won’t switch anytime soon

in blr after almost a year. will be here for 4-5 days ;)

i see lot of flaws in this approach how maintaining a continuously compressed memory inside the context window work in real-world cases? this is basically optimizing for achieveing sota in benchmark say if the compression policy go wrong, or the observer misjudged importance, how to recover the original context? keeping everything in a context means the model always pays the cognitive cost of reading the whole observation block. even if compressed, 40k tokens of observations is not less. for agents running continuously for weeks, that grows heavy. and moreover why to spend so much unnecessary tokens in the first place when retrieval systems can handle those nuances better.

people are eventually catching up on RLM

very cool implementation around self-coordinated agents. as nico mentioned, the agents broadcast the files they plan to work on upfront, which helps reduce wasted parallel work. the next step from here is to push this further with scoped reservations, auto-resolution when agents collide, and better evals around whether this actually reduces conflicts. this is similar to what claude code released yesterday with agent teams but open-sourced and more explicit. love this community!

during our user calls we also came to similar conclusion. this is a big pain for lots of teams who move fast. the specstory team has been putting a lot of thoughts and work around the same to have a properly designed flow to have better provenance and a way to map code to prompts and intent directly but in a digestable way without overwhelming the reviewers. let's see.

we are living in a x bubble i believe a lot of my friends at big product companies and mncs (as swe/sde) have never even heard of codex, haven’t tried claude code, and agents like droid or opencode aren’t even in question. most haven't tried those tools personally either because for them spending 20$ out of their pocket for a work tool is not worthwhile. even where copilot or some unheard ide tool is allowed, it is usually used as an autocomplete or "ask a question" assistant. very few large firms have adapted to running long-lived coding agents, giving access to the agents execute tests, run scripts, or modify repos autonomously. it will atleast take more 1-2 years for the firms to accept the reality and figure out how to adopt the changes in the domain while addressing security and data concerns.

my analysis in the past 1 year has been currently consultants, agency owners, and course makers are printing money and probably even more than product startups or so.

was discussing this internally with the team how each agent has their own folder directory for skills and how it was becoming super hard to handle great to see such efforts taken by openai team around standardization and seems like @opencode too adopted this quickly. kudos!