lmeyerov

28 ms·

421.
by lmeyerov 3y ago
With GDS, we get to skip the CPU and go straight to parallel arrays of network/ssd cards, where 100GB/s is very doable We don't have a lot of our customers doing that yet... but they're definitely some of the more fun on
422.
by lmeyerov 3y ago
A few bits we've learned from building louie over the last year+: * Training on the internet means fault tolerance out-of-the-box * Rule-of-thumb: Sub-gpt4 models are derpy, while gpt4/claude are reasonable. The dumber the model,
423.
by lmeyerov 3y ago
Graphistry / Louie.AI | MULTIPLE: Senior Security Research; Marketing Video Production; Senior Security Sales Engineer | US+REMOTE | FULL-TIME + CONTRACT | https://www.graphistry.com/careers Graphistry is a Gartner-aw
424.
by lmeyerov 3y ago
Expiring certs strikes again! I'd recommend as part of the post mortem to move their install script off their marketing site or putting in some other fallback so marketing site activity is unrelated to customer operations critical path
425.
by lmeyerov 3y ago
Any sense of comparison to Dremio, which helped steward the Arrow ecosystem for doing this kind of thing? (The idea is great fwiw, I've been following them one-off for years, and we have to do elements of these things in how we build l
426.
by lmeyerov 3y ago
I'd actually love to chat from an adjacent/overlapping area for what we're doing in louie.ai in llm-powered investigation & automation. (We're less interested in yet-another-phantom/tines/xsoar, more intere
427.
by lmeyerov 3y ago
If I was any of these companies, I'd totally invest many billions in ecosystem here. Tensorflow (Google) and pytorch (Facebook) are great examples, it can work. Otherwise, hw companies will continue to lose relevance in the growing ser
428.
by lmeyerov 3y ago
Agreed. If there is a niche that is at the intersection of multiple specialties, and it includes GPU acceleration, there is a good chance it is ripe for a startup to get an early mover advantage. Eg, real-time foundation models for audio ar
429.
by lmeyerov 3y ago
I've worked in one of the top computing labs, with top GPU computing startups, have investor money from Nvidia, wrote CUDA for years, and hire people to do write GPU code. And would say, most people -- even Nvidia employees and our own
430.
by lmeyerov 3y ago
this seems super misleading Amazon nailed high revenue growth from the very beginning, just reinvesting in growth & deferring the margin story. They could have stopped at any time. Google nailed high traffic from the beginning, so ad sa
431.
by lmeyerov 3y ago
This is the heart of most real generative AI systems for reasoning about text: index data using this basic technique (chunked document embeddings), and when talking to the AI, the AI looks up documents from these clusters and loads them in
432.
by lmeyerov 3y ago
We are working with enterprise-scale schemas, and beyond SQLs like databricks, are doing Splunk, Elastic, graph DBs, etc, where many tables/cols is common. Part of the trick includes ideas like using embedding models and continuous le
433.
by lmeyerov 3y ago
For AI projects, afaict, 12k stars or forks is more akin to downloads than contributors & downstreams. GitHub is the app distribution, not just source distribution. I've been curious how to model this better..
434.
by lmeyerov 3y ago
Agreed. We have been building our real-time GPT flows for news & social as part of Louie.AI, think monitoring & and investigations... long-term, continuous training will become amazing, but for the next couple of years, most of our
435.
by lmeyerov 3y ago
Curious why? This is maybe 1/3 of my use of GPT4. Quite often, the log dump and nearby code is enough, often even without explicit instructions. Being able to do this task is similar to GitHub CoPilot code autocomplete working well too
436.
by lmeyerov 3y ago
yeah that's beyond my paygrade :) At least for our work, it just takes 1 hat out of the ring for those that we would be otherwise paying attention to, but afaict doesn't really change much otherwise. There are parallel efforts acr
437.
by lmeyerov 3y ago
There are immediate cybercrime etc activities happening around AI -- eg, content abuse is rampant rn -- so someone advising on active standards that explicitly doesn't (relatively) care about these, especially at the federal standards
438.
by lmeyerov 3y ago
Yes es/os has been especially great to the louie.ai team. We need to deal with a lot of data - we handle talking to real-time news, logs, social media, alerts, threat feeds, etc - and these systems were previously already scaled for la
439.
by lmeyerov 3y ago
Datalog <> graph is classic, the points-to analysis papers were some of my first "aha"s for why it's practical If you enjoy this kind of thinking, we recently released GFQL for dataframe-native accelerated graph queryin
440.
by lmeyerov 3y ago
Yes: That's essentially their fine-tuning offerings. They rewrite some weights in the top layers based on your input, and save+serve that for you. It sounds like you would like a wrapped version tuned just for big context. (As others w
441.
by lmeyerov 3y ago
Indeed! You may enjoy this talk where I do just that... end-to-end on GPUs, and < 100loc Python: https://www.youtube.com/watch?v=8ZMzsTbfImU Your intuition about mapping to kernels is good. Basically all SQL, Polars, Duc
442.
by lmeyerov 3y ago
The nice thing about a GPU soln (ex: python dataframes in cudf, just a few loc) is these generally come down to your IO bandwidth, like a single 2GB/s SSD to a 16-32 GB/s PCIe to 1-2 GPUs running crazy fast. And then buy more chea
443.
by lmeyerov 3y ago
Something I was hoping to see is more about APIs for folks making their own specialized experiences. NOAA is doing a good job at raw data, but a lot of gov services consumption & utilization is more specific. We are building some emerge
444.
by lmeyerov 3y ago
Graphistry / Louie.AI | Senior security SE + community | USA REMOTE | graphistry.com/careers Graphistry is a Gartner-awarded visual graph AI startup used by cyber/fraud/BI/supplychain/misinfo/... analysts
445.
by lmeyerov 3y ago
Seeing an entire subnet or fleet is fascinating. I hadn't seen the perf engineer side before, which is interesting. Process graphs are so fun! We get involved in a lot of graph scenarios, and this one we frequently see in the security
446.
by lmeyerov 3y ago
This seems similar to recursive summary indexing and then a choice of which linked chunks to load upon a match There are a lot of 'policy' choices here, so I've been curious how to automatically decide. The KG, GNN, and IR l
447.
by lmeyerov 3y ago
Curious how does this compare to mixtral moe? I couldn't tell from the writeup
448.
by lmeyerov 3y ago
yep!
449.
by lmeyerov 3y ago
I've found altair/vega to not 'feel' declarative as whenever I sit down and try to pass in a json config for basic scenarios... something always gets in the way. GG is composable, so this should work in theory, so I&#x27
450.
by lmeyerov 3y ago
Yes, but not the reasons you're thinking - If you have a fixed time budget and increase the GPU memory+compute available, you can directly query a bigger model. Raw models are basically giant lookup functions, and without the extra mem

More ›