Otterly99

That is also the workflow I settled on with agentic AI.

First, I have very small implementation steps and each represent a commit. Second, I write the commits myself because it forces me to read the implementation and understand it.

Dear person,

  I am sorry if this looks out of the blue, but I was trying to answer to an old thread, and HackerNews seems to lock old threads and doesn't allow for private messages (unless I am too dumb to figure it out, which is always a possibility).

Anyway, I wanted to reply to the very thoughtful comment you left my after I inquired about your responsibilities at a co/op(1) and I wanted to tell you that I am very proud of you and that you are the type of "hacker" that I aspire to be :)

I hope this doesn't come as too forward and I wish you anything but the best.

(1) : https://news.ycombinator.com/item?id=48383220

Great website, I am on the way to learn pixel art so the tutorials are welcome.

I always was fascinated by the stages in King of fighters and Metal slug. So much detail.

My girlfriend was homeschooled and she would definitely have preferred to go to regular school (which she attended for high school).

Home schooling sounds good in paper, but as much as school can suck, it is also an irreplaceable opportunity to learn to socialize, which is a much more valluable skill than knowing pre-university curriculum.

They even clearly say "While this underlying RT, found in a jumbo phage, had been identified in previous studies, Claude appears to be the first to notice the system’s defining features."

The least they could do would be to link to the study or name the authors.

I would say most of these look pretty good but most of them miss the important information except maybe 5.

3 has most of the information but some of it is hard to see.

I would say this is a pretty good way to showcase what is wrong with image generation in AI in general: impressive visually but missing the basics (here complete and readable information).

I have been playing around with Google flow generating scenes with animals characters and the fur always look incredible. Same goes for the facial expressions but the generation is comically bad in scene continuity, even in the same scene. It's not uncommon for characters to lose/acquire features or objects from one moment to another (like losing glasses, but they reappear a few seconds later).

For a lack of a better term, I would call this the "golden turd" effect.

I have learned Czech for a year before moving to Czechia (self-learning) then 4 years doing classes.

When I arrived at my first class, I was a bit lost even thought I had 1 year of "experience". I did progress quite fast the two first years though because I had a bit of vocabulary under my belt, but I had to re-learn all the grammar/declinaison rules.

My advice to you would be: if you can afford, take classes for the language you want to learn, at least for the first year.

Also forget about the quick learning tools like Duolingo and such. It makes you feel like you're learning, but from my experience, not much is being retained.

I'm now switching to learning Russian and taking classes for 1 or 2 years.

How did you mix up keyword and semantic search?

Are they separate searches, or some sort of weighted search?

It depends on what you called solved.

If the goal is to only extract the unstructured text from the document, it is definitely solved. Extracting a more natural structure like paragraph separation, tables, header, footers (what is referred as document intelligence) is much more complicated and not fully solved, but I would say almost.

Weird, I tried it and Jev gave me 16-18% for each face of the die.

You're right sorry, I meant as accurate (in terms of performance, not speed). I would have expected giving access to more elaborate functions to an agent would give better results, but it turns out 4 tools are enough.

A little tip for LLM-as-classifier that worked for me.

Use binary questions rather than multi-class. Then you can use the consistency as another signal of your pipeline working or not, on top of accuracy.

The most astonishing thing to me is that Pi harness is basically as efficient as the Codex/Claude.

I wonder if the same is true for the smaller models in the 9-32B range? I would expect that these models need more steering, but again I was not expecting this result either.

Always exciting to see people working on novel models, rather than the Nth version of the same slightly tweaked LLM.

I'm very curious how much ressources are needed to run such a model. This could be a complete game changer for local applications.

That is also the conclusion that I got from these events.

Unfortunately, it seems that the general response is basically "throw even more RL at it". I'm not sure if it is even possible to decouple the idea of "learning" with reward/punishment systems and loss functions.

The posts came from many different agents, each with a different name, so it would have required roughly the same amount of work.

I never used Sora but I recently tried Google Flow and the results are quite good. I have the feeling that Nano Banana 2 has been the image generation champion since it's release so maybe OpenAI feels that they cannot outcompete Google on this task?

I have been having the opposite problem with my papers lately.

Instead of glossing over mistakes, the reviewer report missing information which is clearly stated in the text.

This is quite frustrating and hard to stay polite after the fifth "please just read the text."

I really wish there was some serious rework of the publishing system. Nobody is happy about it but nothing seems to be moving.

Amazing work!

This reminds me of this youtube channel[1] that I used to watch a lot during COVID where the author animates miniature scenes with their hamster.

[1] https://www.youtube.com/@SIMI_STUDIO/videos

Can you explain how raising an animal can be carbon negative?

It just doesn't seem to make sense to me. How can any mammal capture more carbon than they release?

Thanks for the paper, it is actually quite crazy to think about when SSRI are basically the standard to treat depression.

I recently read a book about "The brain energy theory of mental illnesses"[1] which encapsulates depression and found it pretty compelling.

It seems to be a relatively new and active area of research[2], so there is hope!

[1]: https://books.google.fr/books/about/Brain_Energy.html?id=GIx... [2]: https://pubmed.ncbi.nlm.nih.gov/38183680/

Here is my two sides argument concerning AI in sciences (from a former biophysicist).

Positive side: in sciences, you need to have a lot of transverse skills such as presentation, data curation and data analysis. In biology in particular, students are never taught how to properly store data and even less how to automate the analysis. The pattern I saw was always the same: the data is stored in a convenient way for humans to read, but impossible to process via a script (think patches of columns in separate excel sheets, sometimes the labels are on the side of the value instead of on top). I also was guilty of this at the start of my PhD and had to learn it. With coding agent, you don't have that coding barrier anymore and can plot your data and even get coaching on how to properly store your data so I think this is a huge boon. Second positive side is that you can get insight on papers which are a bit outside of your field and extract some information.

Negative side: AI also allows you to push more garbage, especially if you are working in a less experiment-heavy field, so it will get harder to sift through good and bad papers because now you can make a bad paper look good. This is honestly more of a deeper problem with sciences in general that has the unfortunate publish-or-die mindset where grants are given based on how many publications you have rather than the quality and impact of your publications. This will only amplify the problem as now you can spit out papers faster, meaning the competition can spit out papers faster...

Given how many people eventually quit after their PhD or first post-doc (that's me), I find it hard to argue that we need more scientists in the current academic environnement.

I would even go one step further: Meta, knowing exactly what they were doing, ended building an engine that maximizes ad-revenue.

It's really hard to disentangle Meta's actions with the overall push for maximum-profit-at-any-cost mentality of most societies we live in.

Althought I agree with the first point of the author that FTS is underrated in this new RAG-first framework, the whole article really hides all the problems with RAG-pipeline and kind of hand wave everything.

If you are building a RAG pipeline for your company and are struggling like me, I would recommend this author that has whole series on entreprise documents (start with the one from May 22nd): https://towardsdatascience.com/author/angela.shi/page/4/

Note: I am not the author, just got her article in my newsletter and found it useful.

Super cool story!

I actually had to look up "Come bet on Aaron’s life with us" to make sure it wasn't real.

I also liked your WhatWeSee art exhibition, pretty interesting concept.

I was talking about the waistline drop that the OP mentionned (85.17 -> 84.61).

I wouldn't put my hopes up that peppermint oil has anything to do with it.

That's a shame, because the study in itself is interesting and raises some questions, but the whole time I was wondering why protein translation accuracy would be related to intelligence? Turns out there is no evidence of link and that the title just speculates.

I also was thinking of this film, however I found it really dull.

What did you like about it?

I'm not trying to diss you, I usually like contemplative movies but this one was just too repetitive for my taste.

Very likely. If they really wanted to make the results a bit stronger, they could at least have switched the two groups after 20 days.

It is a 0.7% drop, so very likely to not be significant unfortunately.