mnkv

fundamentally, "linguistic illegibility" is a new term for something that we've known about for about a decade now. In RL the more general ideas is "reward hacking" and in NLP it has been called "semantic drift".

I dislike this term because it doesn't explain where this "illegibility" is coming from. Models are post-trained towards non-linguistic goals with (mostly) non-linguistic rewards. A model's reasoning chain is reinforced if it leads to a correct answer or agentic goal. It doesn't need to be linguistically accurate and meanings can drift over training.

Sounds nice except that these are 1 very small scale model, 1 reranker, and 1 embedding model that are far from frontier LLM level. And they're not open sourced.

As much as I agree with the message, this reads like marketing copy trying to make a big deal out of a tiny model being hosted privately.

you mean like the thinkpad trackpoint?

Just talk to them as if they were already your friend. Most of what you talk about with friends isn't just mutual interests and you start conversations with them all the time.

This blog post describes the basic work of a research engineer and nothing more. The amount of surprise the author has seems to suggest they haven't really worked in ML for very long.

Honestly? This is the best its ever been. Getting stuff to run before huggingface and uv and docker containers with cuda was way worse. Even with full open-source, go try to run a 3+ years old model and codebase. The field just moves very fast.

The paper you're talking about is "Deal or No Deal? End-to-End Learning for Negotiation Dialogues" and it was just AIs drifting away from English. The crazy news article was from Forbes with the title "AI invents its own language so Facebook had to shut it down!" before they changed it after backlash.

Not related to alignment though

https://www.forbes.com/sites/tonybradley/2017/07/31/facebook...

My ideal is copilot that would evaluate the PR against some basic guidelines that maintainers write down.

And perhaps a way to filter PRs to just contributor PRs would be easy to implement and pretty useful

Nice work. A while back, I learned convolutions using similar animations by Vincent Dumoulin and Francesco Visin's gifs

https://github.com/vdumoulin/conv_arithmetic

> the generation of 281,128 augmented examples, from which 1,000 were held out as a benchmark test set.

This model is trained on a custom dataset of 280k examples then tested on 1k very similar examples from the same dataset. Of course it is specialized to outperform general models on this specific task in this specific domain with this specific json format for output.

This is a reasonable hobby project and interesting approach to synthetic data generation but not impressive research.

At minimum you should test your model on other benchmarks that have similar tasks e.g. docbench

reasonable post with a decent analogy explaining on-policy learning, only major thing I take issue with is

> Reinforcement learning is a technical subject—there are whole textbooks written about it.

and then linking to the still wip RLHF book instead of the book on RL: Sutton & Barto.

I think it's pretty obvious it's 1. Given the recent huge, clearly politically-motivated cuts from the current administration, it feels pretty likely that FOIA could be disrupted under the guise of "cost-saving".

And I think you're supposed to be generous to the commenter, not the current administration ;)

how does this compare to zotero?

> IR temperature sensor for checking your body temperature or stuff you baking in the oven

> tiny thermal camera sensor for inspecting leaks in house for the winter

So just a thermometer gun? It costs like $20-30 on amazon and I've never needed one other than in my home / kitchen. Why in the world do you want a phone for this haha.

I do think I've found the perfect car for you: https://tenor.com/view/homer-simpsons-car-gif-8120474

Good summary of some of the main "theoretical" criticism of LLMs but I feel that it's a bit dated and ignores the recent trend of iterative post-training, especially with human feedback. Major chatbots are no doubt being iteratively refined on the feedback from users i.e. interaction feedback, RLHF, RLAIF. So ChatGPT could fall within the sort of "enactive" perspective on language and definitely goes beyond the issues of static datasets and data completeness.

Sidenote: the authors make a mistake when citing Wittgenstein to find similarity between humans and LLMs. Language modelling on a static dataset is mostly not a language game (see Bender and Koller's section on distributional semantics and caveats on learning meaning from "control codes")

The low quality of this blog post (and all your posts) makes me think you wrote it mostly with AI. The irony is palpable.

AI is clearly enhancing humans in their creation of garbage.

You give up a shitty view and any chance of fire escape. Such a savings!

You understand some of what your cats mean because you learned it using the same language games Wittgenstein describes. Also they co-evolved to work with us. But just because you understand three moods of your cat, doesn't mean you would understand their language (if they had one). In fact, it is well studied that cats communicate differently with other cats than with humans.

Sidenote: grouping together non-verbal communication and language fails to take into account the richness of language.

The problem with your perspective is you are assuming children have independent wants and needs to work and can stand up for themselves like adult workers can. This just isn't the case.

Children don't personally decide to work, they are told to by parents / authority / etc.. and they are incredibly vulnerable as employees. Comparing child labour to lemonade stands is ridiculous. There are rules about children working in the family store etc.. and child labour laws explicitly account for this. Child labour laws are explicitly about preventing abusive conditions in factories, fields, and other hard, grueling jobs.

Of all the absolute horseshit in the article, the most obviously wrong is saying that "tech doesn't play both sides of the political aisle".

So obviously wrong if you look at any single company. Just recently, we saw how FTX was lobbying hard for both sides.

And there absolutely has been a backlash against tech. Ask any Facebook employee how people react when they say where they work.

> The solvents we need in this case are the healthier methods of fulfilling these longings.

Terrible take. Algorithmic feeds and ad-tech has continually optimized how to get and maintain our attention. It is ridiculous to think "going into nature" or any individual solution is the answer. We blame pharmaceutical companies for making addictive drugs, why don't we blame tech companies for making addictive apps?

How did you run the benchmarking, zero-shot or few-shot? I think a fair comparison would be Llama-7B which got an average ~35% for 5-shot.

Did you read the whole report??? The complainant is essentially stalking his neighbour with surveillance footage, recording any time the neighbour so much as looks in his direction, and keeps a stalker-ish log book with insults. On top of that his wife and his father nearly hit the neighbour's 4 year olds with a car and then say they refuse to slow down. The report says the car was "inches away" from another person and that kid would have been run over!

And then the complainant has the audacity to file criminal suit over his neighbour giving him middle finger after he does it first! This story is about basic insanity.

The important difference between the LM and the content moderation system (itself built on top of an LM) is their training objective. LM is doing next-word prediction (or human-preference prediction with RLHF), whereas the content moderation is likely finetuned to explicitly identify hate etc...

So while the LM is not supposed to output "truth", the content moderation system should correctly classify "hate" because that is its training objective

Even if you take this perspective, different groups are considered vulnerable in different parts of the world and it seems the API uses the American perspective. E.g. Scandinavians are a less protected group than Italians according to the article. But if you are in Italy, you will likely find Italians to be less vulnerable. Even if you agree with the rankings, this just means you come from the same cultural perspective and others will disagree.

This is a fundamental problem with identifying hate / bias and I think its important to understand our current limitations.

They should add geolocation to the fuckups. That is not where either Montreal or Gatineau are haha

The victims are cyclists because so many dutch people cycle! (Because the infrastructure is so good)

Per Capita, cycling in the Netherlands is far far safer than cycling in the US.

This article is interesting but a major issue is that the vast majority of the points (and the linked articles!) are not about psychedelics. They are about pharmacology, psychology, statistical significance, and science in general. The majority of the author's critiques can and are applied to a lot of other fields (he even takes a meta-review critical of nutrition and applies the logic to psychedelics).

His main three themes (author bias, statistical significance, double-blinding/scientific rigour/confounding) are not at all unique to research with psychedelics. If the author's point is that we should take all studies with a grain of salt then why focus on psychedelics? If psychedelics is really such a problem field then I would expect more evidence than a couple bad actors (the MAPS stuff sounds terrible!) and light issues in a few studies that were brought up in reviews before publication.

It seems the main point isn't research as much as the public's perception and hype (partially fueled by researchers to be fair). That's a good point and I'm totally against the psychedelics version of "weed cures cancer" bros. But, as the author points out, this isn't unique to psychedelics either! Ironically, I found the intro and most of the article to sound much more critical of the research than the conclusion really is.

I'm all for basic competency in math. But looking at the specific court ruling (https://www.otffeo.on.ca/en/wp-content/uploads/sites/2/2021/...), I don't think this standardized test is necessarily the best way to get there.

The EQAO (in charge of creating the test) found through a literature review that "the fundamental goal of teacher licensure tests – to improve student learning – is often not met." Adding math courses to the B.Ed would be a much better way of improving teachers' math proficiency. It seems strange to have an extra standardized test when you have a specific program designed to graduate teachers.

I was going to point out how old all of the info in the article was (why talk about Kyoto when the Paris accords and COP are more recent and more relevant) but then realized the article is from 2017.

Ignoring the nonsense idea that climate change would be good for Canada, many criticisms are not true anymore. Keystone XL is cancelled. Justin Trudeau (and the Liberal party) ran their last campaign on a plan to meet the Paris accord targets. Government subsidies are more and more being moved from oil and gas to sustainable energy. There's a whole lot more to be done and lots of fair criticism but I doubt even Jesse Brown would be as strongly critical today.