GPT-4o made a chess engine

[smatovic]

[...]
When I learned to drive I didn't watch 1000000TiB+ videos of training data.
[...]

Well, it took you 16 or 18 years until you were ready for the drivers license

Progress is quick now, I say AGI/ASI within 6 years (2030), if we do not blow the planet otherwise up.

--
Srdja

[towforce]

The whole AGI talk is just snake oil. We don't even have fully auto driving AIs on cars w/o killing everybody on road.

Are you telling me human brain is just a loss function w/ a sum of neurons connected to it? Come on.
When I learned to drive I didn't watch 1000000TiB+ videos of training data.
LLM it's just auto complete on steroids.

It's good for creating templates. But writing good code you need a good human programmer to sense the quality.

1. We do have "fully auto driving AIs on cars w/o killing everybody on road": Waymo's cars have a better safety record than humans

2. I agree: no human will ever look at anything like the number of chess games used to train today's NN models. We can all agree that they're learning/training in a very different way from humans. I argue that they're picking up a huge number of shallow/simple/surface features of chess, whereas the human way of analysing picks up deep/complex patterns. I believe that even deeper patterns exist than even humans pick up, and I'm going after those.

3. Studies show that LLMs make programmers more productive (a third more iirc). Remember when the best chess was produced by humans and computers working together? That's where we are with programming IMO. Now, the best option is to just let the engines choose the move, and it's likely that this is where we're going with programming.

Having a negative attitude towards today's AI doesn't alter the fact that they are doing a lot of very useful work right now.

[towforce]

[...]
When I learned to drive I didn't watch 1000000TiB+ videos of training data.
[...]

Well, it took you 16 or 18 years until you were ready for the drivers license

Children can drive cars: they're just not allowed to.

I vividly remember when my boy was 9 or 10, sitting in the back of an off-road car with adapted controls while he drove it around a course!

[smatovic]

[...]

[...]

[...]

One for the critics:

Springer-Paper: ChatGPT is bullshit
https://link.springer.com/article/10.10 ... 24-09775-5

...but Bob, it draws nice pictures?

--
Srdja

[towforce]

One for the critics:

Springer-Paper: ChatGPT is bullshit
https://link.springer.com/article/10.10 ... 24-09775-5

That's quite a bit of text to read - thank goodness I've got Gemini to summarise it!

"Large language models (LLMs) like ChatGPT generate text that appears human-like but often contains factual errors. These errors have been termed "hallucinations," suggesting misperception. However, this paper argues that "bullshit" is a more accurate term.

What is Bullshit?

Bullshit, as defined by philosopher Harry Frankfurt, is speech or text produced without concern for truth. This can be:

* Hard bullshit: Intentional deception about the speaker's agenda.
* Soft bullshit: Lack of concern for truth without deceptive intent.

Why ChatGPT Outputs Bullshit

LLMs are designed to produce convincing text, not necessarily truthful information. They prioritize what sounds plausible over factual accuracy. This aligns with the definition of soft bullshit.

The paper further argues that ChatGPT can be considered a hard bullshitter because its primary function is to imitate human speech, potentially deceiving users about its lack of genuine understanding or intentions.

Why It Matters

Describing LLM errors as "hallucinations" is misleading. It implies a misperception of reality, which isn't the case for machines that don't perceive. Instead, "bullshit" accurately reflects their indifference to truth.

This distinction is crucial because how we frame the issue influences our understanding and expectations of AI. Viewing LLMs as bullshitters helps us critically assess their outputs and avoid overreliance on potentially unreliable information.

Conclusion

The paper concludes that "bullshit" is a more apt term than "hallucinations" for describing the false statements generated by LLMs. This terminology better reflects the nature of these errors and promotes a more informed and critical approach to AI-generated content."

[towforce]

Another one for the critics: some employees are starting to use LLMs to produce fake work that's not easy to differentiate from real, valuable work:

[Werewolf]

Claude 3.5 Sonnet has been released in the last few hours. They claim it's better than Claude 3.0 Opus (Opus = their big LLM).

Presumably, 3.5 Opus will come at some point.

In the same way, I believe we are also waiting on an update to "Google Gemini Advanced Ultra 1.0"

[towforce]

Claude 3.5 Sonnet has been released in the last few hours. They claim it's better than Claude 3.0 Opus (Opus = their big LLM).

Presumably, 3.5 Opus will come at some point.

In the same way, I believe we are also waiting on an update to "Google Gemini Advanced Ultra 1.0"

That's good news, and I can see it's more powerful just on a quick try - but I haven't got time to have a good try right now. Good news: it's still better than Gemini at doing exactly what it's told - so I will still use it for some uses (sometimes it's better to let the intelligence shine (I put life coaching in this category), but sometimes you want the job done exactly how you specify it (I put planning my day in this category)).

IIRC Gemini Advanced only came out in February - and they make announcements about improvements to it nearly every week. Why do you think we're due a major new version?

[MaxDemian]

Idk to me it doesnt sound so unrealistic to get a LLM to be really strong at chess. If you have a lot of good training data, for example training data which has as input a chess fen from a GM game and as output the move the GM played and you have like a million or many millions of these training data examples and train/finetune a LLM on it would it really be so unrealistic that it picks up patterns and starts to learn chess and suggest decent moves on new positions? Such training data could be created with some python script for example which loops through a huge collection of high lvl GM games.

[MaxDemian]

And if you manage to create a lot of training examples which include explanations of the chessposition/moves you might even be able to make a LLM which can explain chess decently to humans without saying complete nonsense as most LLMs on the market do when it comes to chess