threatripper

FPGA is not efficient in any way except producing very few pieces of a particular logic very quickly.

But I think the higher level logic could check out. If we can abstract the majority of the compute into dedicated ASIC then we might use something similar to FPGA glue to connect and reconfigure them to adapt to new model updates. A bit similar to LoRA layers that you find tune to adapt the model to your particular needs.

I assume that it starts making sense once you have big multi year contacts to run a particular model with only minor updates.

Tens of millions or billions? Either way that's not a blocker if it promises to pay off.

Earlier there was no AI to catch it.

The longer you wait the faster you will go.

Not sure if it's fine but when it pays it doesn't have to be.

I don't see a hard reason. If it works it works. No hard need for a good, universal, and long lasting solution. At some point you just stack slop on top of slop and it works for your use case - and if it doesn't you'll slop it out yourself.

Based on my own experience AI (Fable, Astra) still has major blind spots and is prone to draw obviously wrong conclusions from thin air. Conclusions based on weak evidence that are so obviously blatantly wrong that you question if it can think at all. At the same time it already completely outperforms humans in many cases in many areas. The utility of current AI is in what you need. Can you take the good parts and cheaply compensate for the weakness OR do the mistakes completely kill it for you?

I would therefore not call it AGI (Artificial General Intelligence) because clearly it isn't but ASI (Artificial Special Intelligence) which clearly it is in many areas.

The problem is that ASI may be enough to win wars but not enough to save the economy.

We need to start numbering these things. I opened the board today and thought "oh, another one again?!" but it's just the thread of yesterday.

It seems to start understanding how bicycles work. Look at the front fork. There are reasons why it's shaped the way it is in real bicycles. If you do physical simulations in 3D with reinforcement learning you will start understanding that most of the arrangements that the other agents use just won't work.

There is plenty of evidence that they have improved in all benchmarks and also in my private experience. But have they improved in the things they still fail at? No, they still fail at them. You need only one example of failure to prove that it still fails. They still fail a lot on many real world tasks.

So, depending on what you ask, they may have not improved even a tiny bit.

Apparently that's necessary and sufficient for getting views. In these interesting times when the future is very uncertain we like to hear people speaking with certainty about the future. Especially people who appear to not be paid for it.

Can he predict the future? No. Can he pretend to be able to predict the future? Yes.

Being early is the same as being wrong. Being lucky is the same as being right.

True, if you have a codebase that works in practice but has dozens of loose ends and poorly defined edge cases than it can chase off into rabbit holes because "oh wait, what if x is undefined instead of null? How is y defined? This outdated package has long known severe security holes and should not be used anymore, do we actually need it?".

Nobody except corporations who built workflows on top of it and don't care about the price because the developer already moved on and nobody wants to touch it.

Figure 1 made me laugh out loud.

I fear that we can only predict them if we deliberately trigger them to release the energy.

1/f noise basically kills averaging. You collect more signal but at the same time equally more noise.

Apparently LLMs are very good at decompilation. I mean it makes sense, they are trained on readable code and they understand logical structure very well. And a lot of real-world problems involve inverse thinking.

https://reveng.ai/blog/training-an-llm-to-decompile-assembly...

If you pay your employees enough to sing all day about how much they love you they will sing.

It's not just speed, it will consume a lot less energy per token, maybe even more than 100x difference. And cost for a chip that runs that one model will also go down a lot once volume scales up. They will end up way cheaper than flexible GPU chips.

I expect AI models chopped up into building blocks where 99.9% of the compute is fixed but glued together with flexible "fine tuning" layers that will adapt them to specific applications. Those kind of chips will run 99% of consumer AI and at some point be integrated into consumer devices.

The real crank you need to look at is where we dig the resources out of the ground. All of that will eventually end up in the atmosphere. Plug that hole. Stop digging and drilling.

The invisible hand of the free market will take care of the rest. If red beef consumes too many resources then it will be too expensive to consume. Simple as that.

Unfortunately fossil fuels make a lot of money and people are hungry for more, so we can't do anything about it really. We can only shift blame and distract to pretend we care.

How many miles does the burger drive my car? Not to question your numbers but they lack context. Also, it would be good for nature if there were fewer or no humans on this planet. But then there would be no human here to enjoy nature.

Yeah, it's way easier to argue that the product is a bad idea if they give you all the resources to build, release and grow it but there's just not enough customers showing up.

The alternative might be internal battles to kill it that take up time and energy and everybody can continue their argument forever.

Do you get promoted for not launching a product? The interests of the individual and the institution do not always perfectly align.

Can we assume that the test is still private when it was run on many cloud providers?

Also add electricity 50% on top for cooling the DC.

Just got cut off mid-task in central Europe.

When I open a new chat it still advertises "Extended through Jul 19"

One theory is that they are removing the limits altogether and the update has gone wrong.

Anthropic is arguably still better in tooling and integrating model and tooling. Good habit beats raw intelligence.

For code editing Cursor editor tooling is even better.

To me the biggest gain I see is that you take the programmers out of the loop. Instead of formulating your ideas to start a project and then acquiring the resources to do a single iteration on it which may take months if not years, now many specialists "just do it" and do several such iterations in a single day. Afterwards they may still go the ordinary route via programmers but on a completely different level and a lot of fruitless work is frontloaded and 100x cheaper. This doesn't show up in any such statistic.

So, the question is not "Does it make our programmers more productive?" but "Does it make our organization faster?".

I have great hope that the CAD & FEM field will benefit greatly from LLM use. To my understanding we currently don't have a free CAD kernel and broadly applicable FEM solver because these are just too hard to make because geometry and physics are hard unforgiving problems. But all pieces of information are available and many implementations covering some details exist. So, the task is a lot of literature research and putting everything together and that's where LLM based agents should shine.

Of course we won't 1-shot a new CAD kernel but the human work will likely shift more towards identifying failure modes, specifying test cases and sometimes solving really hard details. Sometimes even working out details that have no precedent in literature or free code.