jimmySixDOF
Joined 1,726 karma
- jimmySixDOFOne of my soft milestone tests for AGI is if this gets reproduced in a World Model with Gaussian Splats or whatever it is by then that lets you do gameplay walkthrough in first player egocentric 3D-6DOF-360 view in XR with some friends till then its just all stochastic parrots on word calculators whats the point
- The business model at Groq basically morphed over time so the internal cloud was their only client and all purchases were on some revenue sharing basis to finance set up and operate the cloud business. So this has a bigger impact on those Cloud hardware operators to the extent they were involved in the discussions with Nvidia. Saudi Aramco comes to mind, as an early big check investor, who hosts much of the Groq Cloud today. So now Nvidia is their sole source supplier and the whole Tokens-as-a-Service business model they signed up for is re-negotiated ?
- How is this determining a comment is "Highly Upvoted" on HN where that is explicitly kept out of the ui/ux and new comments are a/b tested across users etc... it's not a thing you can know. I have had comments with a big number of descendents but very low upvotes so thats not a reliable indicator either. Genuinely curious.
- This story should be bigger news on HN not maybe just for the aquihire but for the statement of intent market signal the future has Johnny Ive's vibes both OpenAI and Meta have now poached top people from Apple AVP teams + Meta is increasing focus on the Ray Ban formfactor + neither has a mobile phone hardware ecosystem to fall back on as source of compute so next couple of years progress in VLM, DETR, Splats, etc might be shaped by the new interfaces and I am well psyched up to see where this puck is going !!
- Is Tesseract still considered the go to here for OCR I would have thought lot of other options are out there now
- Chunking on hierarchy is a good and async built in and a cross encoder mode .... I like this project's Keep It Simple Stupid approach without skipping on functions even a basic graph triple. Using this to fill out a PoC mockup could be worth it vs dummy data and just drawing a cloud.
- Given 10min of sunlight the body can naturally produce 15,000UI equivalent so I think gp is likely astroturfing for that brand
- You are missing the point where accuracy stays the same
- Reveal.js vs Sli.dev seems like a toss up I am sure there are nuanced differences or maybe I am missing something obvious ?
- Dagger (in Docker) had a idea like this while and Pydantic is using external state savers like Temporal.io there are a lot of directions good luck on yours !!
- I thought for a sec it was insigths.hn revised and back up again I like that you used the Amazon Strands agent framework $2/month is a nice price point I could also see a flat one time X$ + BYO keys approach I signed in through google to test the few threads I have summd look accurate the commenter standout insights check out. I wonder if the agent could setup HN user login and upvote comment etc. Good work pat those agents on the back !
- Interesting to see anync event driven stateless runtimes used here this is becoming a trend for multi/sub agent workflows and using queuing systems in some cases like https://deliveryhero.github.io/asya/ (for kubes) lot of different approaches attempting to deal with fan-in conflict locks who would have thought a swarm of idiot savants would introduce coordination problems lol
- Terminal Bench 2.0 just dropped and a big success factor they stress is the hand crafted phd level rollout tests they picked aprox 80 out of 120 with the incentive that anyone who contributed 3 would get listed as a paper author this resulted in high quality participation equivalent to foundation labs proprietary agentic RL data but it's FOSS.
- AI-RAN is the strategic play here because it's unknown (outside of research lab NDAs ?) what potential real-time physical AI/ML implementation will have on the future of edge processing like organizing the low-layer 6G spectrum contention mechanisms. It's a near certainty that custom AI accelerators are a part of every radio base station in the near future so this is not cash investment but a new product line Joint Venture similar to the Intel story.
- you can read the labels this (-y) uses modernBERT and even has an eval comparison to the (-ie) in it's GitHub so you can see the improvement as tested -- although if you want to do vanilla rules based chinking for whatever reason your data needs then (-ie) is still good.
- I did look into DataRobot's Syftr which points at the same problem but is a lot heavier I definitely like that the approach you take is at least easy to get a basic version up and can start checking the results right away!
- I had my hopes on this project RawDog using local smol sized LLMs but it hasn't been updated in a while. I feel like all this should be running easily in the background nowadays.
- this kind of DSpy-GEPA self improvement loop keeps popping up and adding a few points but the cost (API and wall clock)also means you use this where a repeatable task/prompt/context needs optimizing and you can afford to find better templates
- I liked how Karpathy explained part of this problem as "silent collapse" in his recent Dwarkesh podcast. Meaning the models tend to fall into a local minima situation of using a few output wording templates for a large number of similar questions, and this lack of entropy diversity it becomes a tough hard to detect problem when doing distillation or synthetic data generation in general. These algorithms as nice python functions are also useful repurposed for labeling parts of ontology and topic clusters etc [1]. Will definitely star and keep an eye on the repo !
[1] https://jina.ai/news/submodular-optimization-for-text-select...
- Also why the sudden interest? Amazon Alexa snips have been used before in court/investigation and this is not new. But makes me wonder about what happens when you are dealing with summaries of summaries of long gone tokens. Is that evidence?