rajman187

Cerebras got lucky with the ban on NVIDIA chips due to "security reasons" in the middle east, securing a massive deal with the UAE's sovereign fund. Of course, the leaders made a nice donation to the Trump crypto fund and coincidentally the ban on NVIDIA was lifted. That presented a danger for Cerebras, but they also pivoted almost entirely to an inference as a service company, maximizing tokens/sec metrics on a variety of open weight models. Then they landed the contract with OpenAI which has bought them another lifeline, and helped paper over the trail of fraud that delayed the S1 and IPO for a few years. Fascinating story indeed!

i'd say 99% of usecases will be served fine with Postgres, or a managed instance via AWS

i think it's worth revisiting this in a short while because, by and large, how the engineering craft has been for the last 40+ years is no longer the correct paradigm. it takes Claude Code a few moments to put together an entire proof of concept. engineers, especially experienced ones, will be less likely to produce (and hence be performance-calibrated on) code as output but rather orchestration and productionization of [a fleet of] agents. how do you guide an llm to produce exactly what is needed, based on your understanding of constraints, available libraries, various systems and APIs, etc. to accomplish some business or research goal?

in that sense, estimation should theoretically become a more reasonable endeavor. or maybe not, we just end up back where we are because the llm has produced unusable code or an impossible-to-find bug which delays shipment etc.

Re: cerebras, they filed a S1 [1] last year when attempting to go public. It showed something like a $60M+ loss for the first 6 months of 2024. The IPO didn’t happen because the CEO’s past included some financial missteps and the banks didn’t want to deal with this. At the time the majority of their revenue came from a single source in Abu Dhabi, as well. They did end up benefiting by the slew of open source model releases which enabled them to become inference providers via APIs rather than needing to provide the full stack for training.

[1] https://www.sec.gov/Archives/edgar/data/2021728/000162828024...

> You could cut your MongoDB costs by 100% by not using it ;)

Came here to say exactly this

They’ve filed a S1 [1] last year when attempting to go public. It showed something like a $60M+ loss for the first 6 months of 2024. The IPO didn’t happen because the CEO’s past included some financial missteps and the banks didn’t want to deal with this. At the time the majority of their revenue came from a single source in Abu Dhabi, as well

[1] https://www.sec.gov/Archives/edgar/data/2021728/000162828024...

If I could upvote this more than once I certainly would

My main keyboard has been a 34-key split Ferris. I usually have either a trackpad between the halves if I’m using a Mac or an ergonomic Logitech if on my Linux desktop. Not having to move my hands at all while being able to reach any keys/characters I need has been a welcomed change, worth remapping my brain.

https://arjtala.github.io/2022/09/17/ferris-compact.html

Yeah the org structure is one thing, the missions are another. Yann adds some clarity here

https://www.linkedin.com/posts/yann-lecun_were-excited-to-ha...

This has nothing to do with the newly appointed fellow nor Meta Superintelligence Labs, but rather work from FAIR that would have gone through a lengthy review process before seeing the light of day. Not fun to see the license change in any case

An intuitive treatment of RLHF, TRPO, PPO, GRPO, DPO and RLAIF

That’s why you have encoders as well as decoders. For example, another model from Meta does this for translations; they have encoders and decoders into a single embedding space that represents semantic concepts for each language

https://ai.meta.com/research/publications/sonar-sentence-lev...

Not a lawyer but would assume downloading material from libgen is, in the vast majority of cases, illegal because it's a breach of copyright or similar. That’s gotten Meta in quite a spectacle of late [1]

[1] https://www.loeb.com/en/insights/publications/2023/12/richar...

Well there was the case of an employee leaving due to his perceived moral issues around the use of copyrighted material in the training dataset [1]

[1] https://www.pbs.org/newshour/nation/openai-whistleblower-who...

> other than a bit of open source (PyTorch and React are nice, I guess)

Not to detract from your main point but I think this misses a lot of contributions, eg Cassandra, Hive, Presto, GraphQL, the plethora of publications coming out of FAIR (fundamental AI research) and of course the Llama family of models which have enabled quite a few developments themselves

> It was clearly valuable from day 1

I’m not sure that’s the case even if in retrospect we can clearly argue this

In 1998, Paul Krugman, winner of the Nobel memorial prize in economic sciences, infamously predicted that “the growth of the Internet will slow drastically, as the flaw in ‘Metcalfe’s law’—which states that the number of potential connections in a network is proportional to the square of the number of participants—becomes apparent: most people have nothing to say to each other! By 2005 or so, it will become clear that the Internet’s impact on the economy has been no greater than the fax machine’s” [1] and even though that turned out to be spectacularly wrong it shows the attitude in the early days wasn’t one of absolute certainty.

And certainly the current AI boom is most visibly known for LLMs but there is a lot more happening beyond chatbots.

[1] https://www.snopes.com/fact-check/paul-krugman-internets-eff...

It originates in Yann LeCunn’s paper from 2022 [1], the term AMI being district from AGI. However, the A has changed over the past few years from autonomous to advanced and even augmented, depending on context

[1] https://openreview.net/pdf?id=BZ5a1r-kVsf

From the documentation [1]

> The mission of Sail is to unify stream processing, batch processing, and compute-intensive (AI) workloads. Currently, Sail features a drop-in replacement for Spark SQL and the Spark DataFrame API in single-process settings.

[1] https://docs.lakesail.com/sail/latest/

MTIA will be for inference initially. Another to add to the list is wafer maker Cerebras

https://www.forbes.com/sites/craigsmith/2024/08/27/cerebras-...

Meta is already working on this [1], not sure it can replace NVIDIA for training large models within that time frame however. The ecosystem around their chips is what gives a huge competitive advantage, not having to build entire libraries and optimize a ton of code goes a long way in adoptions and retention.

N.B.: there are other 3rd-party competitors like Cerebras [2] who offer all-in-one solutions for their giant wafers along with libraries and data centers but I’m not sure behemoths would migrate to these offerings either

[1] https://ai.meta.com/blog/next-generation-meta-training-infer...

[2] https://www.cerebras.net/

The title and opening is perhaps giving people reason to infer something which Lecun isn’t arguing, that AI isn’t going to reach such a level of intelligence. Indeed, he’s published on the topic of Autonomous/Augmented Machine Intelligence [1] and while a subtle difference between that and AGI, it’s not negating the possibility. Just perhaps not through existing LLM architectures

[1] https://openreview.net/pdf?id=BZ5a1r-kVsf

I would have liked to see a more generic implementation that isn't necessarily tied to NVIDIA, while I agree that's a much greater ask than a 7-person team can probably take on, there's a whole cohort of ML and data science folks working on, say, MacOS that are missing out on this optimization.

In a world of finite time and resources, wouldn't it be more useful to improve services on the lines themselves? The Elizabeth line had the highest rate of cancelations in the entire UK between July and September

https://www.standard.co.uk/news/transport/elizabeth-line-can...

https://www.theguardian.com/uk-news/2023/dec/08/elizabeth-li...

While I agree the public transit in the US is abysmal at best, I'm not sure the UK is a good measure anymore.

Rail strikes and engineering work are happening with such frequency that getting around seems to take longer and longer. A 1+ hour commute to just travel 10 miles is pretty absurd.

Insane cost overages and years of delays aside, the Elizabeth line has suffered from nearly 10% of all services being canceled in August.

At one point I was needing to take the Northern line from Clapham and would have to wait 3-4 trains before I could squeeze on. The max I waited was 10. Then I moved away.

Add to that infrastructure issues like Hammersmith and Wandsworth bridge closures and commuting becomes a soul crushing prospect.

I don't know what the solution is but with such a high tax rate in the country you'd expect much better. Other European countries seem to have it figured out.

https://www.bbc.com/news/uk-england-london-66754481

Several years ago Walmart dramatically sped up their online store's performance by storing images as blobs in their distributed Cassandra cluster.

https://medium.com/walmartglobaltech/building-object-store-s...

The point you’re missing, and with no fault of your own, is that this isn’t the end goal by any means. You’re right that this type of device will always be a niche market, indeed Apple are targeting just 1M units sold.

If you dig deep enough, you’ll find that the ubiquitous, always-on, socially acceptable form factor is indeed AR in the shape of glasses. As in Google Glasses but with infinitely more utility. The ability to do anything you can now on your phone and more, powered by a localized and shared augmented experience.

> What would be such indicators?

If you agree on the above as the end goal, one clear indicator of success is “how many times did you not take out your phone?”

Of course any expert in the field will tell you we’re 10+ years away from achieving such technology, due in no small part to the physics of optics and displays. But you can imagine the Apple VisionPro shrinking toward this end state.

You’re comparing a service accessible through any device with a physical product, not very apt I’d say

Carmack is no doubt brilliant and it shows through Oculus as a product. But AGI is not about algorithm optimization the way 3D graphics were in the 90s. I am not sure the approach he takes (let’s find a way to write this in assembly) can apply here. He’s said before that AGI is going to be a few thousands of lines of code, ultimately [1], and it’s hard to believe that’s really the case.

That said, I did get a chance to speak with him one on one last year and he really emphasized a few things; the need for being product oriented and giving customers what they want rather than chasing cool engineering (über)solutions; not being careless with resources just because we have more power with modern hardware (he poked fun at React where you spin up a new thread just for an interactive button); and being aware of the inefficiencies brought on by infinite resources (# of engineers and/or funding, which make you think less critically about timelines and delivering within bounded means)

[1] https://youtu.be/I845O57ZSy4

The author of Bazel came over to FB and wrote Buck from memory. In Google it’s called Blaze. Buck2 is a rewrite in rust and gets rid of the JVM dependence, so it builds projects faster but it’s slow to build buck2 itself (Rust compilation)

The author of Bazel came over to FB and wrote Buck from memory. In Google it’s called Blaze. Buck2 is a rewrite in rust and gets rid of the JVM dependence, so it builds projects faster but it’s slow to build buck2 itself (Rust compilation)