nsagent

As it can run in the background doing “chores” while I sleep.

Slightly off topic, but this reminds me of the night family in Rick and Morty.

Sadly, the major ML/AI/NLP conferences are being inundated with AI slop papers. That will arguably have a bigger impact on the quality of research moving forward.

  Wouldn't it be wiser to get out of the market into fixed rate assets like government bonds?

I just hope the companies aren't considered too big to fail. Bailing them out would be a bad idea.

https://www.openmarketsinstitute.org/publications/no-bailout...

Sadly, like intelligence and reasoning, I think emergence is a word that people discuss while having different understandings of what the word means.

In the paper I linked, they talk of a particular take on the word that is commonly discussed in the ML research: between one discrete step and the next a phenomenon came into existence without existing before (i.e. it occurred spontaneously). The paper then goes on to refute specific claims. That's all I'm pointing out (I could point out similar issues with research into reasoning).

I think it's difficult to pin down "emergence" in the general sense as it likely goes into a deep philosophical argument like consciousness, so I'm unwilling to state whether or not some notion of "emergence" actually takes place. In that sense I agree with you.

Then they are naive. I've seen this a lot first hand during my PhD. Interpreting results in a way that reinforces their beliefs and ignores alternate hypotheses.

Frequently, those alternative hypotheses are demonstrated to be true. For example, "emergence" as was researched and proclaimed by many labs, including places like OpenAI and Anthropic is due to using too few discrete steps in sampling a continuous phenomenon. See "Are Emergent Abilities of Large Language Models a Mirage?" [1] for example.

[1] https://arxiv.org/abs/2304.15004

Hate to bring the whole Bun rewrite drama into this discussion, but it seems the Bun team thinks it makes sense to trade a potentially faster compile time in Zig to a likely slower compile time in Rust due to the claimed advantages of a memory-safe language with a robust type system.

That approach seems entirely counter to the claims in this post. Obviously, different teams can have different perspectives, but I'm surprised just how different they are (especially for someone on the Haskell board).

I'd love to better understand why the LLM flow works for Bun, but feels untenable for Scarf. Maybe the Scarf team could learn from the Bun team's approach?

I'm sorry, but I'm not sure what works best for a beginner. I started my PhD when the original Transformers paper [1] had just been released. I had no background in NLP whatsoever and used the original paper to write a Transformer and the full training pipeline from scratch during the first couple months of my PhD without referencing any existing code (only reading the paper and it's references).

So I'd say, if you're motivated you could do the same. That said, I've always been a self-starter and I started my PhD after working for a decade. I'm sure there are other resources out there, but I'm not equipped to say what's best for a beginner (I found the original paper to be excellent, but most everyone during my PhD, including my advisor, found it to be inscrutable; I think it's written more like an engineering focused paper, which might be why researchers found it difficult to grok, but with my previous industry experience it seemed quite clear).

[1] https://arxiv.org/abs/1706.03762

I'm unwilling to speculate whether or not OpenAI is breaking their agreements (I honestly have no clue), but as an NLP researcher I'm certain they could launder data by having an LLM rewrite it and subsequently train on the rewritten data.

Papers like "Curated Synthetic Data Doesn't Have to Collapse" [1] and "How to Synthesize Text Data without Model Collapse?" [2] demonstrate it's possible to do this.

Since OpenAI's Privacy Policy [3] explicitly allows for the use of deidentified data, it's possible they consider rewrites (maybe paired with a model used to identify explicit PII) to be deidentified. Whether OpenAI's legal team thinks rewriting in this way technically means they aren't training on your data isn't something I'm able to comment on.

Here's the relevant Privacy Policy statement:

  We also aggregate or de-identify Personal Data so that it no longer identifies you and use this information for the purposes described above, such as to analyze the way our Services are being used, to improve and add features to them, and to conduct research. We will maintain and use de-identified information in de-identified form and not attempt to reidentify the information, unless required by law.

[1] https://arxiv.org/abs/2605.07724

[2] https://arxiv.org/abs/2412.14689

[3] https://openai.com/policies/privacy-policy/

The restaurants get to set those values though. A coffee shop near me puts the default options as $1, $2, $3 for low order totals. So if I go in and get a coffee (which is $4.50), the lowest tip option is 22% and the highest is a whopping 66%.

I've seriously gotten tip fatigue and have been working to move back to a sane standard. I've noticed the places that have these crazy tips, also pay their staff well.

I was actually surprised to hear my brother-in-law deride LLMs as being useless for areas he has expertise in when I visited for the 4th of July.

He was complaining that he would ask how to perform a certain repair on a car, and the LLMs he tried (ChatGPT & Grok) would give him a long involved process and he'd ask why not do it this simpler way and it would say, oh you're right! He just found it gave bad advice and realized (rightly) that in areas he has less expertise in he has no way to judge how good the outputs are.

This is from a guy who loves tech, historically worshipped Elon, loves his Tesla, and (rightfully again) didn't buy into SpaceX because he thought it was overvalued.

In the past when I visited for holidays he was liable to have a positive outlook on LLMs and their utility. Seems telling that he's starting to see the cracks.

MacPorts

Your links show you used a fuzzer, but that doesn't address the other half of Andrew's statement. Is Andrew misreporting/misremembering your conversations?

EDIT: It's really telling that asking a factual clarification question is somehow downvote worthy. I probably shouldn't be surprised, but this epitomizes the reason online discussions devolve in to flame wars (even moreso than real life, though it happens more and more there as well).

The answer could be as simple as we didn't use a fuzzer until recently so both are accurate. I honestly don't know, which is why I'm asking. Yet somehow just asking is triggering to people.

Treat it as a data point and make your own decisions (he does back it up with several sources that indicate a workaholic mentality with that expectation for others). Hiding it ends up perpetuating the behavior unseen.

*Below is an aside that explains why I think it's better to air these things even if it seems like "rumors and gossip."

I know first hand how toxic some people are irl compared to their public persona. There was a professor in my department during my PhD who was known to be a slave driver, but there are no accounts of it outside of the department. We would have to warn incoming students about working with her, though sometimes they wouldn't believe it could be as bad as we said. I spoke with one of her students after she left the lab due to being hospitalized for exhaustion from overwork: the professor contacted her while she was still hospitalized and asked her to complete a task ffs!

Andrew could have been more tactful in his blog post (though as I've grown older I often find less tact to be more effective), but it really sounds like accommodating Bun was a net negative for Zig.

  When Jarred joined the Zig community about 5 years ago, I described him as someone who had strong "beginner energy".

I think that post highlights a difference in philosophical approach to software development. Bun is focused on moving fast and will deal with the consequences as they come. Zig takes a more planful approach.

I've experienced both in my career and I fall solidly on the planful side. It's why stopped using homebrew and I've avoided huggingface packages as much as possible.

Not surprising. It seems that the comment section of every coding agent thread has at least one person mentioning they use "tokenmaxxing" to increase their token usage because it was brought up during their quarterly review, at a standup, or some other communique from on high.

Just wonder what happens when more and more companies introduce similar restrictions. Will that lead to devaluations of the LLM companies?

From the article he linked:

  If a subsequent probe fails, the system cross-references threat intelligence platforms like Shodan to match the host and falls back to their specific indexing methods.

This is the fundamental tension in law making and government in general.

Leaving room for nuance reduces the seeming capriciousness seen in the enforcement of some laws that look heavy-handed when applied strictly, while said underspecification can allow for abuse instead.

As long as people are individuals with their own volition this tension will exist.

You appear to be arguing with a strawman; at least you seem to have ignored a central part of my message:

  The issue is that the test is positively correlated with success in an undergraduate program, so they threw out the baby with the bathwster. The real issue is that the SAT is not able to distinguish the capabilities among students to the degree it purports to.

Having made graphics using TikZ, it was tedious and time consuming, but the results look great! Would have loved an easier way to make them.

There are three major parts of the modern GRE: Verbal, Quantitative, and Analytical Writing. You could easily look that up, or ask if you didn't know.

Responding off the cuff without any reflection on the comment you're responding to doesn't move the conversation forward in any meaningful way. It just comes across as disrespectful.

Do you also think LLM leaderboards accurately reflect the capabilities of the models being tested? If you do, then I can easily point you to numerous academic papers pointing out the various flaws in many leaderboards (from poorly designed benchmarks like bABI and the original SQuAD, to data contamination, and more).

In that same way, any test, including the SAT and GRE have flaws. They can be gamed in ways similar to LLM leadeboards: test prep makes you better at them. That's one of the main reasons universities moved away from SAT; they were afraid that it disenfranchised lower socioeconomic status students (and it does to some degree). The issue is that the test is positively correlated with success in an undergraduate program, so they threw out the baby with the bathwster. The real issue is that the SAT is not able to distinguish the capabilities among students to the degree it purports to.

And if you want an anecdote to match all yours, the first time I took a GRE practice test, I got a 3 on the writing. Not because I'm poor at writing, but because I didn't really know what they were looking for. After reading a test prep book, I got a 4.5 on my next practice test and a 5 on my final practice test. When I finally took the actual GRE, I got 6 on the analytical writing. Trust me, nothing changed in my writing ability over that time. In fact, I didn't even practice the skill except through those three practice tests. Clearly the test was not capable of determining my real ability to make an argument; it merely tested my ability to adapt my writing to what was supposedly being tested.

Interestingly, the vast majority of universities that got rid of the GRE requirements for PhD programs are not going back on that. Turns out that the students with the highest GRE scores are the ones most likely to drop out of their STEM PhD. [1]

[1] https://journals.plos.org/plosone/article?id=10.1371/journal...

I moved from custom POSIX shell scripts to chezmoi due to the complexity of handling lots of random environments. It really simplified getting up and running on disparate server environments.

The last few years I've rarely had to switch machines, so maybe chezmoi is overkill now, but it works well, so I'm in no rush to simplify again just for the sake of it (and who knows if I'll need chezmoi's features for real again in the future).

Get the new Powerbeats Pro 2. Nearly identical in functionality to Airpods, but they have ear hooks as they are designed for sports.

Having seen plagiarism first hand, sometimes it exceedingly blatant. Like copying from a PDF that was produced via LaTeX — since LaTeX hyphenates words to split them across lines, if you end up keep-ing the hyphenation in, the te-xt reads like this.

Yeah, it's been known for a very long time. Richard Feynman alluded to it in his speech The Value of Science [1] where he discussed a Buddhist proverb:

  To every man is given the key to the gates of heaven; the same key opens the gates of hell.

  What, then, is the value of the key to heaven? It is true that if we lack clear instructions that determine which is the gate to heaven and which is the gate to hell, the key may be a dangerous object to use. But the key obviously has value: how can we enter heaven without it?

I know this isn't going to be a popular take, but here goes anyway...

The complaints that Anthropic are routing your requests to a different model reminds me of an old Louis CK bit about airplane wifi. Clearly Anthropic was too aggressive with whatever guardrails they put in, but the response seems overly entitled to a model people didn't even know existed not that long ago.

https://youtube.com/watch?v=me4BZBsHwZs

Having lived in bumblefuck Alaska for a year, I can honestly say that they do in fact pay more, but it's also super expensive to live in rural Alaska.

Likely a bigger issue is that very few people want to live in a town of 3000 people or less that isn't connected to the interstate road system. Money can only do so much to fix that.

Am I reading this correctly? Their chosen cloud providers run the PCC stack on their hardware, so the compute provider is responsible for ensuring the privacy guarantees? I assume that would add to the potential security surface area.

They really are trying to convince the skeptics about AI privacy.

Do they allow you to opt out of data collection to improve their models for Siri? What about allow users to choose on-device only processing?

If not, they are only speaking to the converted when they have Craig drill home their supposed privacy guarantees.