I think there are many other significant inference improvements which an be built out that are being ignored because the API between client and inference stack would be tricky to nail down.

This is not a criticism of the research but instead the presentation but the repo looks very fishy, it isn't clear that this is from a bunch of researchers from Berkley. It also does not make very clear (on the GitHub) what optimizations or performance they are targeting.

I think engineers building very complex systems now have a lot of performance knobs to twiddle that would have just been too costly for human effort. Since we constrain the responsibilities of the agent slop is less of a problem. We relegate it to defined tasks with clear API boundaries and test harnesses.

That's what I keep saying about the van Goghs I burn to heat my home but everyone is still mad at me!

> On a motorcycle, you’re in the environment instead of sealed off from it. Lavender above Grasse in June, a 5-degree temperature drop when the road cuts through a forest, leaning into corners with your whole body

Very romantic, but not very realistic. Most commuters need to show up for work in heat, rain, snow, etc.

I'm hoping to make my next car an electric car but all of the US ones are massive.

[0] - https://www.mitsubishicars.com/mirage-history

1. Selling access to the US? 2. Ship more and better software? 3. Talk about it publicly?

If you had a secret AI better than anything currently available the mere mention of this would crater the US tech investment sentiment.

They could even host these Chinese models through AWS and sell at-cost inference on Bedrock and obliterate the US model companies.

Undermining out entire economy by subsidizing the release of DIY versions of our main economic drive sounds like a huge win for China.

AMD would love to come by and plop one of these into your datacenters: https://www.amd.com/en/products/accelerators/instinct/mi400..... Also, in a few generation, there will be a massive glut of used hardware.

Unless frontier labs can surpass the current models significantly I don't know why I would pay them money instead of hosting K3.

It's amazing how complicated we made things with bad software engineering and how good we can make things with good software engineering.

I often hear that people are "uninterested in what AI has to offer" but I don't really understand it. I understand this is a mix of ethics and engineering that I don't often engage with but I do not understand why someone would explicitly state that they are not interested in improvements to the world based on the source of the improvement.

The comment I am replying to did not say that LLMs cannot create improvements. That would be a different discussion! It instead says that it does not care about fruits born from the LLM tree.

It seems highly likely to me that through a combination of LLMs and traditional optimization software like vizier that manufacturing, industrial processes, scientific research, or even bakeries will see some improved output through the mixing of domain specific expertise and enablement from LLMs.

I think right now, in the extreme hype environment we live in where people say AI will do everything and take over the world and take your job and give you nothing, people are burned out. Once we get past this point through I am hoping people start talking about what LLMs can do that will actually improve our lives and which can actually work. Much of what people tell me LLMs can do, don't actually work reliably. But there are many cases where LLMs have been a step function in capabilities.

Examples: OCR that really works, fuzzy search, fixing scoped bugs in software, building prototypes of software, etc.

Cement board + roll on or something like kerdi system.

https://youtu.be/B5DVpruLygw?si=jD8eMHHGQoWG4hI2

https://openai.com/supply/co-lab/work-louder/ - $230

https://www.elgato.com/us/en/p/stream-deck - $130, with LCD sceens, works with any apps

So, you could think of it like this:

New customer: 100k-500k. Existing customers: 5% rate hike or 1M-100M/year in increased payments.

Per customer this may be minor but there are many customers.

The problem has been we have no system for proactive construction of new generation. Note: up until recently in the US demand has been mostly flat. We have been building significantly more power efficient systems so utilities did not know we would see a massive demand spike.

China somehow was able to harden against this and continued to improve infrastructure.

I don't know if this is the end state, but I think this is pretty close to what I was seeing a year ago and what we will be seeing next year. Yea, maybe we use agents to do the looking through but you will still be shepherding that.