lmeyerov

31 ms·

361.
by lmeyerov 2y ago
The rest is neat but scary for most production scenarios, while a prompt autotuner can give significant lift + resilience in a predictable & maintainable way to most typical LLM apps Again... I'm truly happy and supportive that aca
362.
by lmeyerov 2y ago
He is referring to laws like SB1047 that will make open source model makers liable for who uses them, and may scare companies like Meta from releasing any more. E.g., preventing the release of llama 3 400b, which might be the first GPT4-gra
363.
by lmeyerov 2y ago
We tried dspy and a couple others like it. They're neat and I'm happy those teams are experimenting with these frameworks. At the same time, they try to do "too much" by taking over the control flow of your code and runn
364.
by lmeyerov 2y ago
Yeah I would recommend sticking to RAG on naively chunked data for weekend projects by 1 person. Likewise, a consumer tool like perplexity's search engine where you minimize spend per user task or go bankrupt, same thing, do the cheap
365.
by lmeyerov 2y ago
Entity resolution - RAG often mixes vector & symbolic queries, and ER improves reverse indexing, which is a starting point for a lot of the symbolic ones Identifying misinfo - Ranking & summarization based on internet data should be
366.
by lmeyerov 2y ago
Agreed, model training and data collection are great! The subtle bit is just doesn't have to be for LLMs, as these are typically part of a system-of-models. E.g., we <3 RAG, and GNNs for improving your KG is fascinating. Likewise, d
367.
by lmeyerov 2y ago
We work in some pretty serious domains and try to stay away from fine tuning: - Most of our accuracy ROI is from agentic loops over top models, and dynamic RAG example injection goes far here that the relative lift of adding fine-tuning isn
368.
by lmeyerov 2y ago
The last point is spot on... Pandas on GPUs (cudf) gets you both the perf + usability, without having to deal with issues common to stack/array languages (k) and lazy languages (dask, polars). My flow is pandas -> cudf -> dask cu
369.
by lmeyerov 2y ago
Related: If you are into this kind of thing and the extra fun that is GPUs + pydata, we have a 1mo or 2mo project around adding rootless to our GPU graph AI containers & packer flow. Ping build@graphistry.com . Niche but a project we ha
370.
by lmeyerov 2y ago
I find SQL nice for a happy path - highly structured, non-nested, and simple log scenarios, such as some heavily curated app log subset like an app-level authentication audit log When we have arbitrary logs from all over, and especially sys
371.
by lmeyerov 2y ago
In our experience building louie.ai for a continuous learning variant of text2query (and for popular DBs beyond SQL), getting syntax right via a symbolic lint phase is a nice speedup, but not the a correctness issue. For syntax, bigger LLMs
372.
by lmeyerov 2y ago
To keep control in the hands of the analyst, we've been working on UX's over agentic neurosymbolic RAG in louie.ai -- Ex: "search for login alerts from the morning, and if none, expand to the full day" That requires gene
373.
by lmeyerov 2y ago
Yep, I very occasionally see natural '10x' types who can jump into any codebase or business.... and more often see '10x' types who learn the environment more slowly but then speedrun using their other skills once they ar
374.
by lmeyerov 2y ago
Yes, imagine data analysis scenarios like Excel users or Jupyter notebooks, or operational investigations like user 360's and security incidents. Just now defaulting to natural language and connected to your data silos and a variety of
375.
by lmeyerov 2y ago
Bend is intriguing -- 1. Some potentially useful perspectives: * Weak scaling vs strong scaling: https://www.kth.se/blogs/pdc/2018/11/scalability-strong-and-... ? * ... Strong scaling, especially compari
376.
by lmeyerov 2y ago
we do multistep programs in louie.ai via a variety of agents/tools, like "get X data from DB Y, wrangle cols A+B in Python, and then draw an interactive map + graph" The ultimate answer is fairly short if you are a senior pyt
377.
by lmeyerov 2y ago
I read that comment a bit differently: - Volume: Former founders who hit the jackpot are investing in a bunch of startups every week. There are hundreds of unicorns etc out there, and thousands of Series A's every year. That's a b
378.
by lmeyerov 2y ago
I thought the test limericks were autogenerated?
379.
by lmeyerov 2y ago
That's true today, no reason to be true tomorrow, else we would all still be on mosaic or netscape or firefox (who fired the rust/servo team afaict once they weren't as valued)
380.
by lmeyerov 2y ago
Their GitHub does ;-) I enjoyed the article. We are doing a more genAI-oriented notebook env variant for louie.ai aimed at database/API users preferring natural-language-first vs typical Jupyter/observable/hex/etc python
381.
by lmeyerov 2y ago
TailScale made key parts of their stack proprietary, route key bits centrally, and I'm sure other issues. And, just like with browsers, alts are good. I'd expect a lot of whitespace for a more open, secure, and usable impl. Most p
382.
by lmeyerov 2y ago
Long context windows get confused, so shorter is better, and they cannot fit everything in general. I'm not sure where you are seeing results that say otherwise. RAG is effectively prompt context optimization, so categorically rejectin
383.
by lmeyerov 2y ago
Afaict gorilla, as in that thread ;-) Nexusflow probably too, as it also does function calling and would need to bake in, or explicit fine-tuning for RAG use, which I don't recall seeing I haven't look recently, but there is also
384.
by lmeyerov 2y ago
We mostly use via pygraphistry, and the demo folders have a bunch of examples close to what we do in the wild: https://github.com/graphistry/pygraphistry Ex: ``` import graphistry graphistry.nodes(alerts_df).umap().plo
385.
by lmeyerov 2y ago
A took us a bit to want this: We added a heterogeneous dataframe auto vectorizer to our oss lib last year for a few reasons. Imagine writing: `graphistry.nodes(cudf.read_parquet("logs/")).featurize(**optional_cfg).umap().plot
386.
by lmeyerov 2y ago
Yeah we advise Graphistry users to keep GPU umap training sets to < 100k rows, and instead focus on doing careful sampling within that, and multiple models for going beyond that. It'd be more accessible for teams if we could raise t
387.
by lmeyerov 2y ago
A fed official was explicit about intentionally not inviting meta (llama3) or other oss leaders to their AI policy group, just OpenAI/ms/etc, and in an env when we know OpenAI started with a heavy DC & policy lobbying team sin
388.
by lmeyerov 2y ago
""" This will not apply to a small developer for years. At the point that it does, yes, if you make a GPT-5-level model from scratch, I think you can owe us some reports """ This is an insane way to legislate,
389.
by lmeyerov 2y ago
We have two different external uptime status monitoring services, which has helped
390.
by lmeyerov 2y ago
`query [key] ` A+++, I missed when this got added, amazing! (this basically solves the sticky shared chatroom websocket problem for us when routing to resources with gravity, in our case, multiuser data science notebooks!)

More ›