jononor

Communication tech/tools enable more people to collaborate. It increases ability for labor that is far away from high value markets to contribute. Same goes for shipping tech wrt physical goods. On the global scale that is empowering the labor class. Any productivity tool that individual laborers can purchase also (and that still needs the worker) is probably good for labor, overall.

The one true AGI metric!

This seems like a viable eval strategy. Presumably finding a bug requires some degree of understanding of the code, beyond just information retrieval. However it probably does not measure things like prompt adherence or ability to create code that implements a specification?

Dynamic routing is the usual name for the piece that orchestrates which LLM will be used, based on query complexity. There in an open source implementation as part of the vLLM project (and probably others), it is a field of active research in several universities and labs. It is also suspected that the frontier LLM providers might already be doing something like it behind the covers.

Having local AI as a credible threat will keep them on their toes. Which will benefit consumers a lot.

Hardware sales would be an excellent business model for open weights. Nvidia is already on it with their Nemotron models. Any new LLM/NPU hardware companies would want to so the same, if noone else does it for them (Chinese labs currently do).

Selling managed self-hosting solutions would be another. That is the business of that recent American company.

Selling fine-tuning services or similar adaptations is another. That is what Unsloth is going for, I believe.

Most likely any sound business strategy is going to be of "commoditize your compliments" type. There are many complementary products to open-weight - some probably not invented/discovered yet.

Super info, thanks!

Where/how did you buy your DDR4 from SZ? Interested in doing the same, but want reputable source/supplier.

That is not at all the intention of the ARC team. By ARC teams definition, passing any single ARC-AGI benchmark does not mean that AGI has been achieved. Instead, AGI would be considered achieved when we are no longer able to come up with new benchmarks that the AI systems do not immediately do well on.

Partitioning is not all that expensive. It is definitely worth testing for your specific workload. We use TimescaleDB, which relies heavily on postgres partitions, have a bit under 100 million rows in our active set (last 90 days), across 120 partitions (device*time), and it works nicely. Over 100 partitions is probably a bit many for this workload, but since it works OK we have not changed it.

Network effect means it will be a huge and risky undertaking, and one needs to solve the bootstrap problem. But the costs of video delivery means that one would have to burn serious cash in the meantime. So it works in tandem.

TikTok kinda did manage to make a dent though - I suspect it substitutes for YouTube in some cases (though not all).

YouTubes biggest moat the last 10 years is probably more that all the viewers and creators are already there. Any competitor has a huge disadvantage - creators are not interested in a place without viewers, and viewers not in a place without creators/content.

With enough tokens, all bugs are shallow? :D

Yep that is annoying. There are USB-C magnetic charge adapters. It will prevent shit from getting into the slot, and easy to charge magsafe style. And of course you can easily take it out temporarily to use a standard USBC charging cable.

Time to migrate off Atlassian, and ban it for any use in the company. You cannot just help yourself to customer data like that. The data is not yours, never was, and never will be. Pay for a service that blatantly rips of our company IP? Nope.

Thanks for showing your colors so clearly Atlassian. Good riddance.

A decent amount of software developers and gamers do spend 3000 USD on a PC. That kind of hardware is going go get more and more capable over time wrt genAI models.

Of course there will always be a gap to frontier closed hosted models. It is not an either or proposition.

You are the one that claimed the prices of those shuttle services were lower than that of WaferSpace. 7k USD for 1k chips of 20mm2 at 180nm. Is it not the case?

I went though the list of prices at Europractice. Waferspace is 7000 USD for 1k of 20mm2. That is a per mm2 price of 350 USD. I could not find any offering at Europrice that matches that?

Interesting! Which EDA tool must I use for those, and what is the price of that? Will these services accept a single run of 1k?

What are your thoughts on including a RISC-V hardcore along with the gates? Because for almost all projects I can imagine using FPGA for, I would want a microcontroller also. This might however be slightly colored by me being a software/firmware first type of electronics engineer. Thinking especially for the smaller gate counts, like under 10k - because there a soft CPU takes up very precious resources.

X-ray would be the traditional approach, though quite expensive. IRIS by bunnie is another approach that aims to bring cost way down. Ref https://www.bunniestudios.com/blog/2024/iris-infra-red-in-si...

Nice specs! Looking forward to seeing how this and the other projects on Waferspace goes. Being able to produce 1k chips at a reasonable price will hopefully do wonders for open hardware / open silicon.

Glad to see this. At least there is one player in Europe which does a full vertical integration around LLM/AI - from datacenter to LLM models and applications (Mistral Vibe). On the data center part Europe seems to be doing OK, and also ok on applications. It would be nice to see more players focusing on the LLM model building - though it legitimately seems like a very tough (maybe even bad) business to be in.

Have been playing with Qwen3.5 35B. Runs OK nicely on a RTX5060Ti, though I would have liked to have a bit higher thoughput (a 5080/5090 would do). It is seemingly close-but-not-quite-there for code generation / agentic coding. So I am actually quite hopeful that in a few years time, using local LLM models will be quite feasible.

Very much looking forward to play with the BIO functionality on the Baochips that I have ordered. Thanks for the nice write up! It is fascinating to see how widely applicable the "just throw a RISC-V core or 4 in there" design pattern is. The wide range of CPU designs that are standardized, the number oc mature open source implementations, and the lack of royalty fees, and the ready-to-run programming toolchains really drives this to a new level. And CPUs are small in die area anyway compared to SRAM! Was cool to see on the RPI2350 how they just threw in another two RISC-V cores next to the ARMs.

For these reasons specified above, I think that this trend will continue. For example, in my specialization of edge machine learning, we are seeing MEMS sensors that integrate user programmable DSP+ML+CPU right there on the sensor chip.

Highly recommend Statistical Rethinking for anyone looking for practical/applied/intuitive approach to Bayesian Statistics. For example the 2023 lecture series: https://youtu.be/FdnMWdICdRs?is=KycmwPL-cn8clOK5

Many "subjective" tasks can also be done in an "objective" manner - as long as there is a large enough dataset to estimate what humans would evaluate the outputs - and the evaluators being reasonably consistent. Many human preferences are relatively homogeneous, or sometimes clustered into groups. And there are whole fields of study/practice of such phenomena, such as sensory science - with applications in food, audio, images etc.

Number of machines under control is a measureable target. Quite suited for this concept, at least in theory.

Very nice development towards more open hardware. And super powerful at that. The verification technique is alos super interesting. Have ordered a small set to play with, and to support the project.

There are also several RISC-V microcontrollers on run 1 of Wafer Space, hopefully some of those will also be available online soon. https://github.com/wafer-space/ws-run1

There are easy fixes to get rid of violent and crazy people. Why would a powerful ASI bother with fixing them? A rabid dog just gets put down by humans. Why would we expect anything better of our overlords?