ma2rten

My personal "oh shit" moment was in 2015, when this paper came out: https://arxiv.org/abs/1506.05869

It showed me that a model trained only on movie subtitles data exhibited some (very primitive) reasoning. I have been working on Deep Learning and later LLMs ever since.

Erm, ... OpenAI has hyped when it started and it took 6 years to take off. It's way to early to declare the SSI and Thinking Machines have failed.

Delaying doesn't necessarily mean they stop working on it. Also it might be a question of compute resource allocation as well.

You can add Show HN to the title for your own projects. They will show up in the show tab.

Europe is quite conservative, in the sense that they would not invest billions into an unproven venture. It makes sense that it would excel at an industry that requires putting safety above everything.

It's actually true on many levels, if you think about is needed for generating syntactically and grammatically correct sentences, coherent text and working code.

Interpretability research has found that Autoregressive LLMs also plan ahead what they are going to say.

Your use of the phrase makes no sense. It's the "no parking" that proofs the rule and not the exception.

You can also look at the price of opensource models on openrouter, which are a fraction of the cost of closed source models. This is a market that is heavily commoditized, so I would expect it reflect the true cost with a small margin.

Presumably the model is trained in post-training to produce a response to a prompt, but not to reproduce the prompt itself. So if you prompt it with an empty prompt it's going to be out of distribution.

The study seemed not very convincing to me, at least the way it was described in the article. To summarize: they asked crowdworkers to write a law who used legalese, but not when writing news stories about it or when explaining the law. From that the researchers concluded that people use legalese to convey authority.

But what if people just imitated the writing style of existing laws, but not with the intention to make it authoritative but because that is what they understood their task to be?

This is the same problem as echo cancellation on calls. This is something that built into a lot of software and hardware.

t5x was used to train PaLM 1.

I have an upcoming trip to Europe, which I am quite excited about. I wanted to set up a Tailscale exit node to ensure that critical apps I depend on, such as banking portals continue working from outside the country.

I've never had an issue accessing banking portals from Europe.

Apples cares about the privacy and security of iPhones as a differentiator.

Noam.

No this is not correct. Arguably OpenAI invented LLMs with GPT3 and the preceding scaling laws paper. I worked on LAMDA, it came after GPT4 and was not as capable. Google did invent the transformer, but all the authors of the paper have left since.

Both Amazon and Google already do this, there are reports that Microsoft does as well.

Yes, I think that is a reasonable way to think about it, in my opinion. However, with the language modeling objective it predicts the next token and because of the residual connections each intermediate layer is in the same space. So, maybe it would be more accurate to say that it is an increasingly accurate representation of the next token.

Attention takes in all tokens in the sequence and outputs a new representation of the current token in context. Each layer of the transformer adds more context to the token.

I haven't read this explanation in detail and although they have some nice animations, I wouldn't go to FT to explain machine learning concepts. Here are two well known explanations that might be better:

http://jalammar.github.io/illustrated-transformer/

http://nlp.seas.harvard.edu/annotated-transformer/.

I didn't have time to read this, but it is a single author paper, the author is not affiliated with a research group, it is not peer reviewed, it was published on a preprint server that I have never heard of.

LLMs can definitely perform some kinds of reasoning. For example GSM8K is a dataset of grade school math problem requiring reasoning that LLMs are typically evaluated at. We talk about one method for this in our chain of thought paper [1]

[1] https://arxiv.org/abs/2201.11903

How so?

Someone just asked GPT-4 and got the same result as DeepMind did:

https://twitter.com/DimitrisPapail/status/166684395282416846...

That is only relevant for serving and not for inference, unless the model is too big to fit on a single host (typically 8 GPUs).

https://arxiv.org/abs/2305.15717

It seems very ignorant to bet against AI given all the progress that have been made in that last year (!).

https://jalammar.github.io/illustrated-stable-diffusion/

It's below minimum wage in California.

As someone working in this area, I think this is possible but unlikely. Most likely the model had access to the context, but "forgot" to use it.

It's still useful, but you need to know how to use it.