numba888

I wonder how Meta trains its models. On vanilla Pytorch or they actually have some closed tools and frameworks?

> the unwillingness to put up with shit cooked up by bunch of 25 year old CEOs, CTOs, and the like

You haven't seen the shit written by 50++ senior and principal 'developers'. Imagine local variables names 20-40 chars long. Then they are passed into function call, like 15-20 of them. While actually these are just 3 structures. I.e. 'gurus' unroll them into individual elements and pass. Long names to make the code what they call 'self-documented'. No comments at all. And all this is in a big project with other devs working on it for years. It's absurdly slow for the project of this size and resources. But with almost no competitors this can last for decades and it actually does.

The thing is you may guess and get lucky, but those who control it a lucky almost always, as they know upfront. For example before sensitive announcement politician or other 'big boss' can buy or sell. Through controlled agents to hide his interest, of course.

Trump made some announcements then congratulate BTC holders on record high price. That was his payback and hint for their donations. Smart people got it. But it would be hard for him to invest taxpayers money into this scam. First he knows it. Second he would have to explain and he is not a dictator yet to force it through. As for post-election BTC buyers Trump doesn't own them anything. They are going to loose a lot.

> First of all, it's going to take us 10 years to figure out how to use LLM's to their full productive potential.

LLMs will be gone in 10 years. At least in form we know with direct access. Everything moves so fast that there is no reason to think nothing better is coming.

BTW, what we've learned so far about LLMs will be outdated as well. Just me thinking. Like with 'thinking' models prev generation can be used to create dataset for the next one. It could be that we can find a way to convert trained LLM into something more efficient and flexible. Some sort of a graph probably. Which can be embedded into mobile robot's brain. Another way is 'just' to upgrade the hardware. But that is slow and has its limits.

Does it make sense to use vector search for code? It's more for vague texts. In the code relevant parts can be found by exact name match. (in most cases. both methods aren't exclusive)

I don't think you can generate images with claude. just asked it for pink elephant: "I can't generate images directly, but I can create an SVG representation of a pink elephant for you." And it did it :)

This was nice. I passed it jseessort algorithm. If you remember discussed here recently. Claude 3.7 generated C++ code. Non-working. But in few steps it gave extensive test, then fix. It looks to be working after a couple of minutes. It's 5-6 times slower than std::sort. Result is better than I've got from o3-mini-hard. Not fair comparison actually as prompting was different.

This is probably illegal in China (?)

> I didn't put any full-time SWEs out of work, but I did take one job away from a Fiverr freelancer.

Who would use LLM anyway these days. Interesting when Fiverr will add non-human freelancers. Something similar to algorithmic traders. Passive income.

not sure about Claude but my main problem with 03-mini is that it 'forgets' the things which are supposed to fit in the context window. This results in it using different function names, data structures. I think it's guessing them instead of fetching from the previous records.

Not sure what they found. Either model is unable, or they were unable to solve the tasks using models. Loos like they used strait questions and not Chain of Thoughts. The result for the same model depends on how you ask. The tasks probably required more thinking under the hood than model is allowed to do in one request. More interesting would be if model is capable of solving given enough time. Using multiple requests orchestrated by some framework automatically.

I'm not sure it was even human.

Don't worry, you will get speaking skills with practice quickly. I think you can talk to OpenAI app in Romanian. Not sure. After not speaking for 3/4 of my life I understand only about one half of audio news and bit more of text.

> You may have to euthanize a chicken, likely by hand (literally) via cervical dislocation. It still ranks among the worst things I've ever had to do in my life.

Traditionally it's done by decapitation. Head dies instantly. No need to suffer. Body runs for a while. Don't forget to ask for forgiveness...

duels happened everywhere, that was normal. you can't judge ancestors by modern standards.

you mean this one: “I wholly disapprove of what you say and will defend to the death your right to say...

Stretching and bending can go long way. Did you know there are zillons of sexes now? It's not the brain, it's the wrong body. For each psychological disorder.

The power of brainwashing. The results are somewhat unreal. Britons want Ukrainians to fight more than Ukrainians themselves. The explanation is Britons don't see hundreds of thousands deaths. Their media reprints reports from Zelinkiy office which state just a few thousands. This is called 'Support Ukraine' and they are proud of it. Till the last Ukrainian. Some politicians are more open and clearly say it's a cheap way to kill Russians. They want the war to continue. This is called 'strong position' or 'position of power'.

Some allow it even on live interview. There is no reason to be shy if everybody is 'cheating'.

> scale of the problem in the USA signals an imminent collapse

In my reality so far so good. Finally something is done about national debt which is about $100,000 per citizen. Impressive. Dems kept throwing money left and right, trying to 'feed the world' and themselves on the way. Sock market is good. Bitcoin goes down as sign of likely Ponzi schema collapse or adjustment. Musk probably sold his $1B. Trump very unlikely will create government fund of bitcoins and other cripto-scum inspite of promises to supporters. They had and still have their chance to sell. He is a showman but not stupid. EU has bigger problems now and coming.

You didn't expect it to do all the job for you on PhD level, did you? You did? Hmm.. ;) They are not there yet but getting closer. Quite a progress for 3 years.

> Given the current direction that country is taking that seems like an overall poor career move

Just wonder what is better? China, Europe, Canada, India.. all have their own problems

It's time to start using AI or you will be always too late...

There are bit more expensive servos with proper encoding. The thing is hobby airplanes don't need high precision as pilot controls looking at the whole plane movement. Now that hobby is killed by regulations. Robotics is coming..

Just got for $40 on amazon cheap arm with servos and bearing at the base. "Robot Arm Kits for Arduino Coding". Not that I need it, just like putting things together. Arduino not included, but I have it already. Let's see if I can use Raspberry Pico instead.

When looking at low cost arms I check how it connects to the mount. If it hangs on just a servo it will be very wobbly. Most mentioned in this thread are. For this price or bit more you can get a way better arm on ebay with big ball bearing on the base for stability. I got a kit for $60 or $80 recently. No servos, full metal. It's easy to add servos and put together, for now I have other arms assembled ;) No steppers because I want them to be cheap and imprecise for ML project.

>what am i missing?

1000e kit does not include steppers and board..

It uses cheap hobby servos which is, to put it mildly, not the same as steppers. You will never get 0.2mm precision with them. They will were out quickly and even quicker under load.

BTW, looks like it doesn't have closed loop control. "0.2mm repetability" (should be repeAtability) is only 'under certain conditions', no load.

Smaller robots tend to have smaller problems. Even little help from the model will make them a lot more capable than they are today.

Did they get the first token out? ;) Just curious, NVidia ported it, and they claim almost 4 tokens/sec on 8xH100 server. At this performance there are much cheaper option.