Machine learning in trading: theory, models, practice and algo-trading - page 3744
You are missing trading opportunities:
- Free trading apps
- Over 8,000 signals for copying
- Economic news for exploring financial markets
Registration
Log in
You agree to website policy and terms of use
If you do not have an account, please register
An interesting article about the latest technical developments at Astra. The gist of it is that they’ve implemented a transformer architecture, just like in a good old RNN.
There’s no increase in intelligence, just a reduction in the cost of running the agents. There are also concerns about security.
In my humble opinion, it’s time to start thinking about compulsory electroconvulsive therapy for the ‘AGI creators’ 😝
An interesting article about the latest technical innovations at Astra
Regarding the achievement of AGI, there is a fairly transparent hint at the ‘opacity’ of the achievement:
Astra achieved 99.9 per cent on ARC-AGI-3 — a figure that’s making the rounds in all the headlines. Wilson draws attention to a footnote: this result was obtained using OpenAI’s own harness, which carries over the opaque state between prompts and compresses the context so that the model can continue where it left off. On the standard ARC Prize harness, the same model scored 62.7 per cent.
Well, the benchmark (ARC) itself raises entirely justified doubts as to whether it can serve as an unambiguous measure of AGI achievement:
Understanding the rules for scoring serves as a measure of intelligence only when the system encounters them for the first time and solves them using logic.
As soon as engineers start feeding the model terabytes of similar tasks and giving it an infinite number of attempts to refine its code, ARC-AGI ceases to measure intelligence (AGI). It becomes a trivial test of graphics card performance and the quality of the algorithm’s fine-tuning for a specific test.
Understanding the rules governing points serves as a measure of intelligence only when the system encounters them for the first time and resolves them using logic.
It seems that’s exactly what happens there. What’s more, the ‘puzzle’ was solved in fewer steps than the average person would have taken.
Aleksey Nikolayev #:
This result was obtained using OpenAI’s own harness, which carries over the internal state between requests and preserves the context so that the model can continue where it left off.
I may be wrong, but it looks as though this is exactly how a human would have done it.
It seems that’s exactly what happens there. What’s more, the ‘puzzle’ was solved in fewer steps than the average person would have taken.
In my humble opinion, it seems obvious that intensive training to pass an IQ test will not increase one’s general intelligence, although it will improve one’s score on the test itself. The same applies to models.
I may be wrong, but it looks as though a person would have done it themselves.
In my opinion, the problem lies in the lack of transparency regarding what exactly the person conducting the test did – what specifically led to such a huge improvement compared to the standard procedure.
As an example of human influence on test results, unrelated to this specific case, there was the Meta benchmarking scandal, following which Yann LeCun left the company. It seems that different models were used for different tests (which most likely means that the model was tuned or selected specifically for a particular test in order to inflate the results).
If I understand correctly, a highly intelligent person (such as Leonard Euler) shouldn’t automatically be good at playing chess. Rather, they can learn to play very effectively.
To learn. By himself. Not by having a group of scientists rewrite the contents of his brain through a complex, opaque procedure.
I’m not sure whether the following is a sign of intelligence.
A person is given the rules of chess. And they are told that they must learn to play well.
A person creates a computer programme (based on artificial intelligence or, as in the 90s) to play chess.
For example, the programme achieves grandmaster level (modest by modern standards).
The person plays chess exclusively through their programme.
Could one say that the person has not learnt to play chess? Probably so. Did they use their intellect to play chess? – I don’t know.
...Did he use his intellect to play chess? – I don’t know.
That’s more specific and easier.
Then you’ll be able to categorise them by type and complexity, and make comparisons.
And only then will you be able to be surprised when you piece it all together and try, first and foremost, to define intelligence for yourself. Without appealing to scientific authorities, government or legal stamps with signatures, and so on.