killerstorm

There are several papers demonstrating AI is at least as good at diagnostics as fully qualified doctors. So why would NP + AI be worse? AI should compensate for the lack of knowledge.

And I'm not saying NP + ChatGPT - it should be properly calibrated system which would defer to a 'proper doctor' in more complex cases.

You need to train data for a BERT-based classifier, and then there's a risk that it will pick up specific biases from the data instead of what you want.

As far as I understand, the idea of Jev is zero-shot or few-shot classifier: it learns a lot of stuff at pre-training, but unlike a classic LLM it doesn't need to learn how to chat, so it can be much smarter at a particular size

Education system was set up in XIX century. It's not clear how much of it is necessary in XXI century.

Back in the day religious books were copied by scribes educated in a monastic tradition. Now printers can print them in a completely godless manner but the result isn't any worse.

It's low because there's no competition. MDs are like medieval guild: once you're in, you're set for life. Restrictive regulations are lobbied by MD associations, which limit competition.

Probably better solution is to upskill nurses + AI to do handle all the simpler tasks like prescribing standard treatments, etc. There's already a concept of mid-level practitioner which can be expanded.

There's basically no need for GP to be a doctor.

> my ideal sandboxing is "prevent writing to anything outside this dir but still allow reading to most things so that I don't have to manually copy things into a container/VM"

That's what Codex does out of the box, and it's not good against malware - i.e. a rogue npm packet (or even just codex after prompt injection) can read your ssh key and send it to the attacker.

Making model to consists of many small modules is inefficient on GPU, especially as routing adds data dependencies, etc, and especially with pytorch (compared to a custom kernel).

The difference might be smaller on a CPU which has limited parallelism.

But it's basically equivalent to a very deep model which might be problematic for training.

True, but I think in the end people were better off with automation. You know a lot of people were suffering even when they were fully employed.

Luddites also complain about disruption, you know

I'd say this social contract which got "broken" never existed in the first place.

Even before AI we heard lots of complaints like "I made a popular open source library which is now used by corps with trillion-dollar market cap and I don't get anything out of it; halp". There was always some kind of a conflict, now the nature of the conflict just changed

> Look like how AI slop has unrealistic physics

> Posts a link to real moon landing footage

I'd delete the article if I was you...

You know, in academia, they sometimes retract articles, even if they believe they are directionally correct

This is, basically, text-to-LoRA with some extra stuff.

I.e. it basically takes text, computes and embedding and makes a LoRA adapter out of this embedding.

Note that it is equivalent to a recurrent module attached to a transformer. Dynamically generated weights (proposed in the article) are computationally equivalent to multiplicative-gating network with fixed weights. Basically just a beefier variant of GLU operating on a slightly larger state.

Here's what Victor wrote about inets in Bend2 (on X):

> interaction combinators still parallelize better than anything else, but the graph overhead prevents us from compiling to maximally efficient assembly. bend2 is basically inets without the overhead. in a way, inets live in it architecturally, but they don't exist at runtime

From what I understand, the main difference between lambda calculus and inets is that in LC you can refer to a binding multiple times for free, i.e. call same closure multiple times, etc. In inets, you can't - they are more like physical wires where each reference costs. You can definitely see inets in Bend design here (from the guide):

> A closure is affine: it can be called at most once, even when everything it captures is Data. Only top-level definitions can be called freely.

So programming in it might be very different from the normal functional programming. Seems like a big limitations. But I guess that's what lets it run without GC, on GPUs, etc.

A lot of information here:

https://gist.github.com/VictorTaelin/77fd5a2a8a4a07e1da6157e...

https://github.com/victortaelin

A complete implementation have been released, how is that not a substantiation?

Academic people might have more trust in a paper which when through a lengthy publication process. But if you think about it, it's not a better proof than a direct access to the thing. It used to be hard to try out software but with modern tech it literally takes minutes...

That's a start-up style marketing: when you make a product you focus on a big vision and positive sides and de-emphasize weaknesses. I'm afraid that's actually 100% Victor's decision to do it this way, and it seems to be working in terms of generating hype: it got ~4k likes on X, which is a lot for a new language.

Regarding substantiation -- they released source code and demos. As far as I understand, the weakness is that proofs are very verbose as there are no strategies. etc. However, they are making a separate service for making these proofs using proprietary technology: https://bend-lang.com/bender

Victor put 5+ years of research into this. You can find many of previous versions (which use different approach, do a different kind of a thing, etc.) on the github. "Bend2" in particular have been in development for 2 years.

Calling this "a random vibecoded project" is rather disrespectful, don't you think?

Regarding the paper, he states it clearly "designed by the human author". That's not at all the same as just asking Fable to write a paper. I mean the important thing is ideas, not the way they are described.

Please tell me how "I'm glad you're having fun vibecoding" is not disrespectful?

I thought that you thought Bend web site is all that is to it and wanted to point to relevant information. But if you think that "having fun vibecoding" is an appropriate thing to say to somebody who spent many years doing research, I don't know what else to say.

Again, as a "proof of research" take a look at : https://github.com/VictorTaelin/Interaction-Type-Theory that's 3 year old, pre-dates Fable, but OMG doesn't look like a paper.

Victor Taelin has been doing interesting PLT research for 10+ years.

I suggest you read his history: https://gist.github.com/VictorTaelin/77fd5a2a8a4a07e1da6157e...

before making slop accusations. Older variant of what became Bend is 5 years old, so definitely not "vibe coded": https://github.com/HigherOrderCO/HVM1

Interesting that you mention LG and Samsung TVs.

Does it happen on Google Pixel phones?

Obviously, the quality of the walled garden depends on the maintainer. Google's quality standards are lower than Apples, but higher than LGs.

Hmm? Malware on iOS is extremely rare.

I remember in Bitcoin community ~10 years ago, standard recommendation was than an iOS wallet was secure enough (I don't recall even a single case where wallet was stolen via malware), but any private keys on Windows were strongly discouraged, as most cases of stolen wallets were on Windows.

I'd say popularity of iPhone shows which way people prefer, but you do you - what prevents you from voting with your wallet and buying a Linux phone?..

Back in early 2000s, pretty much all Windows machines were infested with malware.

Do you want to bring back those glorious days?

Back in the day users didn't really have much valuable and sensitive stuff on their machines and malware was rather benign - just sending spam, not trying to fuck up that specific user. Could be a bit different when it's a smartphone user depends on.

"Language Models are Few-Shot Learners" - 2020, the GPT-3 paper.

It have been demonstrated that in-context learning is a very powerful mechanism. There's no evidence that models of the size of GPT-6 are bad at in-context learning. In fact, ARC-AGI-3 score might indicate they are good at it.

There's no evidence that a bespoke RL environment is required for each new skill - quite likely a good demonstration is sufficient.

This is an absolute nonsense. Any frontier model can implement chess program from scratch - modeling the board, checking legality, etc. If you asked e.g. GPT-6 to get good at chess and gave it a computer, it will get good at chess. That's an actual strategic skill.

Asking GPT to play chess directly using its reasoning only tests its reasoning ability to model chess state. Which it really is NOT optimized for.

This is also true for humans - people who don't have years of chess training can't really tell which moves are legal given an algebraic notation transcript. These people might have good strategic skills in different areas. Chess is just a very, very specific skill

If AI is good at following instructions, desired behavior can be specified in a (system) prompt

I doubt "slowdown" framing was optimal - could frame it as an acceleration towards more practical AI rather than deceleration

27% more efficient than base32 at expense of all kind of weird compat issues with shell scripts, etc. Yikes!

You're confusing ToS guardrails with instruction-following issues and cheating.

If a model fucks up your tests to report a success, it's not alligned.

boys will be boys

"Generalization" is rather unspecific.

Analogy-making points to more specific mechanism: ability to identify features within representations, and modulating associative lookup using those features. I.e. it's not as simple as D(x, y) i.e. distance between embeddings, but something like D(f(x), f(y)) where f projects representation to a specific feature space.

The evidence can be found e.g. in LLM interpretability - people were able to identify features as something concrete. Also in our cognition - we can tell _how_ two things are similar, or find Y similar to Y in context of F.

"Prediction" by itself seems like a black box: if we are trying to predict sensory input, that gives no explanation to our ability to focus on specific details, or explains why some prediction failures are more likely, etc. OTOH if we say that cortex might be 'disassembling' sensory signal into high-level features and is trying to predict those (features themselves might be identified as "something useful for prediction", i.e. something which explains a lot of variance), it's much easier to connect low-level "prediction" to our high-level cognition.

When my daughter was 6 months old, she once pointed at a towel and said "P". Apparently she recognize Winnie the Pooh from a book she saw earlier.

If you think about it, this requires recognizing something common in things which are very different: e.g. book and towel have different texture and color, and you somehow need to separate what's depicted from the background. So that's pretty much innate, core brain function. I mean, any animal with vision can recognize an object from the background - otherwise vision is useless. But for humans (and some animals) this translates to depiction of object on a flat surface very easily.

So, yeah, analogies-all-the-way-down seems plausible. Even object-vs-background and depiction-vs-paper is itself an analogy.