Almost every item in this list (except Karpathy's LLM Wiki) is based around vector embeddings.
Vector embeddings were super-hot a couple of years ago, but I don't think they have sticking power.
The moment you have an agentic tool calling loop the idea of doing a big fuzzy embedding search and hoping you get back relevant results loses attraction. You can give your agent ripgrep and let it figure out how to find the right results all on its own.
The biggest downside of embeddings is that it's very hard to set a threshold score below which you ignore things. If you ask a vector index for 10 results ordered by similarity you'll get 10 results - but results 3-10 might be total junk.
Very cool, never thought of that! "way smaller" is almost an understatement, when it's 50kb :P Neat that it loads in GitHub READMEs as well, which is probably a large reason people use .gif today.
It's a cool tool/platform, but very different. Asciinema tries to make the "multimedia" itself better by making it actual text instead of being video/images, while the CLI command above turns actual text into multimedia supported by platforms already. Both are useful, both have their use cases :)
I have a bunch of opinionated/personal-use binaries like this in my $HOME/bin/, like delete-all-npm, clean-rust-cache, download-youtube-playlist, and get-markdown <url>. It feels good, and I don't need to remember any commands. Sometimes my coding agent can figure out how to call some of those tools too ;))
This isn't the first time this has happened, either. I do not understand how these consultancies - who sell these "reports" for six or seven digit sums - continue to mess this up. It should be excruciatingly embarrassing for them.
I guess nobody ever got fired for paying KPMG and friends for an expensive report that supported their priors.
KPGM et al. are used as political ammo to push through internal changes. Those in power rely on consultancies underlying their decisions (painful redundancies, firings, etc.). Acknowledging that the arguments for these painful decisions was hallucinated will lead to many problems for powerful people, so for now it's best to just try and sweep it all under the rug.
These six-figure reports are produced by underpaid kids in their twenties working 18 hours a day.
The purpose of paying for these reports is for executives to have someone else to blame when their idea doesn't work. It has nothing to do with the correctness of the content.
> These six-figure reports are produced by underpaid kids in their twenties working 18 hours a day.
That's accurate, for the first draft. Similar to big legal firms - subsequent versions are signed-off and passed up (and if revisions request, down) the hierarchy, each stratum with its own billing rate(s).
Which makes me wonder when the hallucinations got added.
It can't have been at any of the big 4, because partners aren't skipping 4+ org-chart layers to look at draft documents written by early-career associates. I have no experience with body shops - if that's where you were.
I would have to disagree, this report in question sounds more like thought leadership dribble rather than a report commissioned by a client with a scope attached.
The purpose of most reports are absolutely for Assurance to decision makers or management and often times, we disagree with management or provide a view that might not favorable. Which just reflects the realities of what we have identified or tested.
As I said, this seems like thought leadership dribble which absolutely even as someone who has worked in Big 4, I think they're pretty average.
The problem is that there's a lot of people running around who believe the polite fictions we tell ourselves about review processes. It's very hard to explain why it doesn't work to have someone manually clean up a sloppy AI draft without discussing the fact, which many people find unacceptable, that manual review can't catch all errors.
I lurk on the teachers subreddit and get shown videos by teachers on TikTok and the impression I get from that algorithmic bubble is that the kids can't read any more - reading comprehension in particular is terrible. Lots of anecdotes of kids who can't read a few paragraphs and then answer questions about what was in them.
My impression of that sub is that it's a lot of people who went through honors classes in a "good" school district, and are now teaching non-honors, potentially in a "bad" school district and are discovering how the other half lives.
Hasan Piker's political project polls extremely high (universal healthcare, abortion access, and more), so actually you could understand American voter politics by reading Hasan's comments amusingly enough.
As of ~8 months ago the quality is most definitely there, for almost every form of programming I've experienced.
If you're working in some vanishingly rare domain then maybe it's not yet, but most coding challenges are very much in the wheelhouse of the current frontier models.
Nobody writes any code any more at all. Nobody even writes Jira tickets anymore. They don't even review code, and I think we're lucky if they even test it. The AI does all of that.
A small group of developers at my company have set up volumes of skill.md and other instructions for the AI to write Jira tickets, then take action on those Jira tickets by writing the code. The AI submits a pull request. Then there's another AI to review the code. They've written the game plan for the AI to do all of this. All the human does now is click "approve" without even reading the PR, and then someone clicks "merge". There's no coding, no critical thinking by a human anymore except for telling the AI what to do... which really anyone at the company could do. I doubt I'll have a job at this company much longer after 8 years employed there.
A lot of people are genuinely stranded if their phone runs out of
battery. How do they pull up a map, or call an Uber, or phone someone to pick them up?
The context was “Everyone does use the internet for everything today” and “most people don't even realize how the internet is running literally everything”. You don’t need the internet for your phone to have a battery charge.
They were providing some illustrative examples: most people rely on the internet for maps. If they don't have internet (because their battery is flat) they can't do xyz.
The argument here is "most people can get by just fine" without access to the internet.
I tried to pick an obvious example to illustrate how that's not true.
The difference is that, prior to everyone having a smart phone, people had backups for if they ran into trouble. They might simply not go somewhere that they might have trouble returning from. They sorted out their travel plans in advance - someone to pick them up from a location at a time. They memorized phone numbers so they could call from a pay phone if they needed to. They carried cash or a cheque book to pay for cabs.
It's pretty easy as an interviewer to spot when a candidate is hedging on a question, and it's the kind of thing that might get discussed in the post-interview debrief.
"Wouldn't give a straight answer on question X" isn't an instant no-hire, but it's not a positive signal.
This doesn't make sense in practice. He hedged so not sure need to look at other factors vs he picked a side and he selected the opposite of what we wanted no-hire or he answered what we wanted small positive signal need to look at other factors.
It was a question of the form trying to figure out how you deal with "X", and he denied X having ever happened, despite that being a core part of his current role.
He was an internal candidate, we were interviewing him to see if we could trust him with more responsibility (more X specifically), since the new role shouldn't cover up X when it happens. The role involved doing X for himself AND for other people.
Similar to the form of "tell me about your biggest weakness" and you responding with "I have no weakness".
I guess it's a bit different when it's an internal candidate.
I've been on the recieving end of clueless folk trying to make me feel small when I've just been looking for a job, so I might be a bit sensitive about it! Sorry for any offense given. Thankfully I'm beyond that craziness now and can just do what I want for work.
instead of telling him out loud, "hey we see you're hedging and applying bullshit interview speak, your answer isn't sufficient", we asking in increasingly obvious but different ways.
It was an internal candidate so it would be awkward to tell him to his face he was floundering.
I'd love to see credible numbers on that. I find it hard to believe that stupid corporate mandates are responsible for more than a small fraction of usage, but without data I have just my own instincts to go on there.
At my employer (megacorp with tens of thousands of employees) daily use is mandated. Our annual bonuses and pay raises for our performance reviews were explicitly tied to this.
It's a retrospective analysis of an assertion made by NYTimes. The original headline wasn't clickbait, just presumptive, and even so, it's a pretty significant publication that spends a lot of time on the HN front page (alongside you, I'll add). I think it's perfectly fair, and nowhere close to a strawman, to deconstruct that claim a year later.
Vector embeddings were super-hot a couple of years ago, but I don't think they have sticking power.
The moment you have an agentic tool calling loop the idea of doing a big fuzzy embedding search and hoping you get back relevant results loses attraction. You can give your agent ripgrep and let it figure out how to find the right results all on its own.
The biggest downside of embeddings is that it's very hard to set a threshold score below which you ignore things. If you ask a vector index for 10 results ordered by similarity you'll get 10 results - but results 3-10 might be total junk.
reply