niyyou

Feels like the most valuable skill to have as a programmer in times of Claude Code is that of carefully reading spec documentation and having an acute sense of critical thinking when reviewing code.

Sounds like a direct quote from some religious ISIS fanatic.

Of course. Here is a corrected version of your text that fixes the grammar and typos while keeping the colloquial tone:

I'd like to offer a less skeptical view on this, contrary to what I've read here so far. LLMs that act (a.k.a. agents) bring a whole lot of new security and privacy issues. If we were already heading toward a privacy dystopia (with trackers, centralized services, etc.), agents could take that to a whole new level.

That's why I can only cheer when I see a different path where agents are run locally (by the way, Hugging Face has already published a couple of spaces demonstrating that). As a plus, because they're small, their environmental footprint will also be smaller (although, admittedly, I can also see the Jevons Paradox possibly happening here too).

I hope that, to whatever extent this is happening, it's not widespread, and anyone committing war crimes is very visibly and publicly tried in court.

Sorry but you’re either dishonest or totally delusional. Tens of thousands of children murdered, nobody even thinks of putting a stop on this, and you say « it’s not widespread ». What?

There has been so so many atrocities, even before Oct 7th committed by both the Israeli army and armed settlers in the West Bank. There is never, NEVER, anyone being held accountable. Snipers shoot kids on the beach? All good. Torture and rape in your prisons? Fine.

Dude, you should fix your society. You are simply heading towards your own destruction, morally, on the side of public sympathy, and merely as people capable of living with other people peacefully.

edit: minor typo.

Sourced articles like these are good and further proof that Israel is a psychopathic state and society. But what's odd is to depict is as "surprising" or even shocking when the same state has been carpet bombing civilian, including women and children for 16 months straight, causing what is estimated at 300 000 deaths, committing every single atrocity or infringement to the international law possible, including targeting medics, journalists, using starvation as a weapon of war, bragging on it on social media, having politicians incite to eradicate the remaining part of the Gaza population, and I could go on and on with nameless atrocities. So just to put things in perspective, this article depicts a horrible incident, but it is entirely in line with the rest of the Israeli policy, and unfortunately pales in comparison with the ongoing large-scale massacre.

It's not even some "forgotten AI technique" (sigh...). It's been a hot topic for the last 5 years. Used a lot with Variational Auto-encoders, etc. Such a bad journalism.

Behind this rhetorically articulated question, there is a simplistic view of content generation and consumption. First off, if you use LLMs, you are _constantly_ reading stuff nobody else bothered to answer you. So, it's not about who speaks, it's about the what. Then, what you deem as written by nobody turns out (not without irony) an average of so many (possibly) good writings and thoughts, gleaned from various corners of the web. In this regard, I particularly like--and want to shout out--the author's clarity on the work's attribution. It's not his. It is Claude's, hence, everybody's (in a sense). Finally, as a mere form of compressed human written productions, LLMs don't have agency over what they generate. So the prompting, the idea, as well as some editorial decisions are still attributed to the person behind this, hence making it unique in its own way. Instead of seeing as a piece that he didn't bother write, I see it as piece he chose to edit summoning the ghosts of every writer who ever put something on internet, which again, he correctly credits (to the limits of knowing what exact data Claude uses... which is another story).

Hey, I really enjoyed reading your post. It's straight to the point, the list is very rich and well illustrated with examples. I have been coding in python for the last 20 years and I nowadays rarely like reading these posts as much as I did for yours. Kudos.

As Sara Hooker discussed in her paper https://www.cell.com/patterns/fulltext/S2666-3899(21)00061-1..., bias goes way beyond data.

Many thanks! For the record, there is another useful tool in the same vein, that packs a repo to be given to an LLM https://github.com/yamadashy/repomix.

Hi Simon, just read your blog post, thanks for the wrap-up. Just curious, what did you use to make Gemini look at all of your codebase, Aider, something else?

Again. This. is. not. Open-source. At best, open-weights. Clickbait 100%.

On the contrary, ArXiv is for pre-prints, i.e. not (yet) peer-reviewed. Off the top of the my head, it was initially used by physicists who often have huge collaborations and long reviewing time. Then the ML community invaded the space later on. This does not mean a peer-reviewed paper cannot go there of course.

In any serious discipline, ranging from philosophy to mathematics, precision is a requirement. Here, "woke" is everything but precise. It's an umbrella term that the right uses with bad faith to discard any form of social struggle or claim for a more egalitarian society, then part of the left took ownership of the term (reverse stigma).

Then a quoted aberrations IMO,

Racism, for example, is a genuine problem. Not a problem on the scale that the woke believe it to be, but a genuine one. I don't think any reasonable person would deny that.

These statements are typically what fuels some people's outrage. Who is PG to decide what is the right scale? In the US for example, too many black people lost their lives because of a systemic racism, at the scale of society (police, job, housing, ...). Is this not scaled enough? To me, it shows an incredible level of disconnection between the social class PG belongs to and the actual problems in society.

It's probably an unrelated post (apologies in advance) but I wanted to shoutout to the Marimo (https://marimo.io), it's the only Jupyter alternative that really got me excited, it's like Streamlit and Jupyter had a kid (and the kid took the best genes from both).

Great comments and insights. Thanks to everybody for being constructive. I am also going through the same situation, but in Europe. Despite having a PhD in Machine Learning, somehow I cannot even get to the interviews. I am suspecting it might stupidly have to do with my LaTeX-generated CV which gets desk-rejected by badly design CV screening systems… anyone else sharing my suspicion?

I almost believed it was a trick to generate labelled data to train AI systems down the line

Precision: « pre-training data is exhausted » everyone has been saying that for a while now. The graph plotting body mass against brain mass… what does it say exactly? (where is the link to the prior point on data?). I think we would all benefit from being more critical here and stop idealizing these figures. I believe they have no more clue that any other average ML researcher on all these questions.

I’ll take the risk of hurting the groupies here. But I have a genuine question: what did you learn from this talk? Like… really… what was new? or potentially useful? or insightful perhaps? I really don’t want to sound bad-mouthed but I‘m sick of these prophetic talks (in this case, the tone was literally prophetic—with sudden high and grandiose pitches—and the content typically religious, full of beliefs and empty statements.

This is a very partial view. A good number of people, including American democrats were disgusted by his unwavering support to the current genocide in Gaza. Many believe he could have stopped it with a phone call and explicitly refused. I know many who swore they would not vote for him, regardless of the circumstances.

Genuine question, is it a property of software design only? Think about construction, for any change in the architecture, one has to demolish stuff and make a mess. I'd argue, that's a general property of change.

So annoying to have articles not point to the original paper.

Worth mentioning another nice initiative https://www.echopen.com/ it’s from Paris and it has been incubated by the hospitals of Paris themselves, first as an open-source project. Disclaimer: I was among the early contributors to the open-source project.

I'm always baffled to see him invited over and over... what is his track record? Since when he is an ML expert? What about holding him accountable for his wrong "predictions" (my favorite https://twitter.com/drjwrae/status/1766803741414699286/photo...)? It is so irritating to see him constantly seeking attention through controversy and calling for debates on twitter.

Sad they do not allow to use a custom OpenAI host address, so one can use a local LLM instead (https://ollama.com/blog/openai-compatibility).

Just try it out before interpreting their vision (which I totally ignored up until now). It is simply a nice system based on your local markdown files, allowing to easily cross-link notes. It is opinionated but I like that. It follows the bullet journaling approach and has a rudimentary integrated todo system.

I have personally tried many alternatives (obsidian among them) but nothing comes close to what logseq offers _to me_, despite its few shortcomings (it's not a lightning fast implementation, it bugs sometimes).

Hope it helps.

Absolutely. Another less thought of aspect is simply power consumption and hence ecology. Running so many remote servers for movings and shows one already have at home is absurd I find.

I truly wonder. Why is Apple doing that? If I'm not mistaken, it is a small niche in terms of revenue.

My hypothesis is that this is perhaps a new approach for marketing: find hobbies that correlate the most with the populations with the biggest potential revenue, taking into account the networking and demographic effects within such a population.

Just throwing this off the top of my head. Bear with me :-).

So many concerns about this one...

- what is the metric exactly?

- what does 90% mean, is it bad or good? That depends on the baseline and the prior distributions.

- on what dataset is the 90% computed? On past crimes? What about sampling bias in this dataset? What if the data focuses on some specific neighborhoods (oh wait...)? What about missing crimes in the data?

In ML, data should fairly well represent a phenomenon and in this case, it does not represent crime occurrence, it only represents crimes detected and documented.

And his visualization of constrained optimization is astonishing https://explained.ai/regularization/index.html (I struggled for a long time to get the right intuition of a Lagrangian)