HNHacker News
TopNewBestAskShowJobs

currymj

3,822 karma · joined March 27, 2016

submissionscomments
currymj··on Will the AI backlash spill into the streets?
software is really weird as a test case for understanding AI automation on white collar work.

it seems exceptionally well-suited to AI-based automation because software engineering has already needed to figure out how to efficiently cope with humans who sometimes produce code which may have defects. most obviously automated testing and type systems. also a lot of programming tasks have verifiable solutions so it's also better for training. it seems natural that the most obviously successful AI tools are for coding.

yet the software industry has also been absorbing waves of automation for 70 years. Fortran and COBOL were referred to as "automated programming" and there was a narrative that these new tools would make it easy for non-specialists to program. this time may be different, or it may just be another wave of automation.

i think software has pretty unique properties among white collar jobs, and I would hesitate to draw conclusions about other industries based on AI progress in software engineering.

currymj··on I got fooled by AI-for-science hype–here's what it taught me
the post that I replied to?
currymj··on I got fooled by AI-for-science hype–here's what it taught me
AlphaFold is real.

To the extent you care about chess and Go as human activities, progress there is real.

there are some other scientific computing problems where AI or neural-network-based methods do appear to be at least part of the actual state-of-the-art (weather forecasting, certain single-molecule quantum chemistry simulations).

i would like the hype of the kind described in the article to be punctured, but this is hard to do if critics make strong absolute claims ("aren't useful in any field whatsoever") which are easily disproven. it hurts credibility.

currymj··on US vs. Google amicus curiae brief of Y Combinator in support of plaintiffs [pdf]
The BERT model, which uses the transformer architecture, was deployed by Google for every English language Google search by the end of 2019.

This is about concurrent with OpenAI's release of GPT-2. But GPT-2 was not really a product.

currymj··on Everyone Is Cheating Their Way Through College
a lot of the ChatGPT assignments are so incredibly obvious. in the most egregious cases literally copy and pasted from the browser without reading. laziness is all-pervasive.

the problem is an academic cheating penalty can have grievous consequences for someone's life, so to prove it requires almost like a "beyond a reasonable doubt" criminal justice standard. and that's difficult to do. so they get through their classes.

"don't hire someone" is a much lower bar. if there are obvious but unprovable ChatGPT vibes in the application, you probably just don't interview them. if there are weird pauses in their conversation during the interview and they sound like they're reading long sentences, don't advance to next round.

a lot of these cheaters are going to get filtered hard.

there is a smarter way to use it but this still requires some level of thought to disguise your cheating, you have to be able to understand the LLM outputs yourself, and a large portion of the cheaters can't manage that.

currymj··on Watching 03 Model Sweat over a Paul Morphy Mate-in-2
> "hey, i am a COMPUTER and a small set of fixed moves should take me about 300ms or so to fully solve out"

from the article:

"3. Attempt to Use Python When pure reasoning was not enough, o3 tried programming its way out of the situation.

“I should probably check using something like a chess engine to confirm.” (tries to import chess module, but fails: “ModuleNotFoundError”).

It wanted to run a simulation, but of course, it had no real chess engine installed."

this strategy failed, but if OpenAI were to add "pip install python-chess" to the environment, it very well might have worked. in any case, the machine did exactly the thing you claim it should have done.

possibly scrolling down to read the full article makes you a rube though.

currymj··on How University Students Use Claude
since late 2024/early 2025 it now is the case, especially with a reasoning model like Sonnet 3.7, DeepSeek-r1, o3, Gemini 2.5, etc., and especially if you upload the textbook, slides, etc alongside the homework to be cheated on.

most normal-difficulty undergraduate assignments are now doable reliably by AI with little to no human oversight. this includes both programming and mathematical problem sets.

for harder problem sets that require some insight, or very unstructured larger-scale programming projects, it wouldn't work so reliably.

but easier homework assignments serve a valid purpose to check understanding, and now they are no longer viable.

currymj··on How University Students Use Claude
Well, a big difference is that arithmetic is something you learn in elementary school, whereas LLMs can do a large fraction of undergraduate-level university assignments.
currymj··on MLB says Yankees' new "torpedo bats" are legal and likely coming
if it leads to more home runs i'm against it.

if it leads to more solid hits and plays in the infield, i'm for it.

unclear which one, i look forward to finding out more.

currymj··on There is no Vibe Engineering
I agree this is true (for mainly political/interpersonal reasons) if you’re writing code inside a typical software engineering organization, which orients itself to producing software products.

But lots of people write computer programs outside such contexts. Then it’s perfectly possible to prototype something and actually rewrite the full version.

currymj··on There is no Vibe Engineering
some responsible yet still "serious" uses would include:

1) throwaway code where the "work product" is not a software system, but rather the outputs of the code which you can verify yourself.

the classic example for me is producing plots. I can easily verify that data is loaded correctly, and that the end result is correct, I just don't want to learn the complex API to make all the ticks and colors and fonts look perfect.

2) prototypes, mockups

3) simple tools (often with a web interface) for your own use

currymj··on The average college student today
the US model has focused less on exams and more on take-home assignments. this can be better because most people perform better without artificial pressure, and it's possible to ask deeper questions than on a timed exam. it's just that it is not viable as of this year.
currymj··on Accessible open textbooks in math-heavy disciplines
i agree with you in spirit (I pronounce Paris as Paris).

however I have never heard of someone pronouncing Freud as Frood, outside of "Bill & Ted's Excellent Adventure".

https://www.youtube.com/watch?v=Sobc2WhL16c

currymj··on What's Happening to Students?
it’s called a flipped classroom and students universally hate it.
currymj··on A Surprise Hair Loss Breakthrough: Sugar Gel Triggers Robust Regrowth
In Mice
currymj··on Goblin.tools: simple, single-task tools to help neurodivergent people with tasks
i can imagine a tool where you take a picture of a cluttered desk or closet, and it produces a list of all the objects and then comes up with a plausible organization plan.

that's a little more complicated than these tools, but it seems within reach given publicly available technology. image segmentation models are really good now, the list of objects would be pretty reliable at least.

currymj··on Tencent's 'Hunyuan-T1'–The First Mamba-Powered Ultra-Large Model
Tone markers are of limited use to Chinese readers (instead, just show them the characters).

They are also of limited use to non-Chinese readers, who don't understand the tone system and probably can't even audibly distinguish tones.

So, it makes sense that we get this weird system even though it's strictly worse.

currymj··on Claude can now search the web
you sort of can do (a decent approximation of) this, it’s just even the approximate version is impractical for computational reasons.

https://www.anthropic.com/news/influence-functions

currymj··on An early look at cryptographic watermarks for AI-generated content
i also doubt practicality. but if the technology worked "well enough" and the major consumer-facing AI companies did it, it would help a lot with certain problems.

if you're technically sophisticated you can easily evade it of course. but 1) most people aren't technically sophisticated, 2) one can ask "why are you going through all this effort to remove the watermark if you aren't trying to deceive anyone"?

currymj··on Deep Learning Is Not So Mysterious or Different
you might enjoy this read, which is an up-to-date document from this year laying out what was the state of the art 20 years ago:

https://web.stanford.edu/~jurafsky/slp3/3.pdf

Essentially you just count every n-gram that's actually in the corpus, and "fill in the blanks" for all the 0s with some simple rules for smoothing out the probability.

currymj··on Deep Learning Is Not So Mysterious or Different
this is pretty close to how language models worked in the 90s-2000s. deep language models -- even GPT 2 -- are much much better. on the other hand, the n-gram language models are "surprisingly good" even for small n.
currymj··on AI Is Making Developers Dumb
it's particularly bad for students who should be trying to learn.

at the same time in my own life, there are tasks that I don't want to do, and certainly don't want to learn anything about, yet have to do.

For example, figuring out a weird edge case combination of flags for a badly designed LaTeX library that I will only ever have to use once. I could try to read the documentation and understand it, but this would take a long time. And, even if it would take no time at all, I literally would prefer not to have this knowledge wasting neurons in my brain.

currymj··on Why Mark in 'Severance' Drives a 1997 Volvo
it's not clear where Severance takes place exactly. But it is wintery and has universities and decrepit fishing towns. So if it's supposed to be New England, i think this makes sense. There are a lot of very successful people in New England who drive old VWs and Subarus.
currymj··on Mathematical Foundations of Reinforcement Learning
generally you are right in spirit.

however multi-armed bandit algorithms are highly useful in practice. these are a special case of RL (RL with one state, essentially).

there are even some extensions of applied bandit algorithms to "true RL", e.g. for recommender systems that want to consider history.

this is the place to look for real-world applications of RL.

also RL uses importance-sampling estimators of the gradient. these sometimes show up in other applications though not framed as "RL".

currymj··on Mathematical Foundations of Reinforcement Learning
Find materials on PPO which should be widespread since it is the most popular RL algorithm. GRPO works on the same principles, just makes certain estimates from samples rather than training an auxiliary neural network to make them.
currymj··on Richard Sutton and Andrew Barto Win 2024 Turing Award
that was “bitter lesson” in action.

for example there are clever ways of rewarding all the steps of a reasoning process to train a network to “think”. but deepseek found these don’t work as well as much simpler yes/no feedback on examples of reasoning.

currymj··on Framework's first desktop is a strange–but unique–mini ITX gaming PC
their marketing photos include 1) a game controller next to the PC, 2) a bottle of Bawls soda next to the PC, so I think so.
currymj··on ChatGPT Can Be Used as Default Safari Search Engine with New Extension
this feature is one where ChatGPT will actually do a Bing search (sometimes rewriting your search query), and then show you a summary of the content of the links.

so "hallucinations" are more rare, although the model may be very gullible with respect to the content of the search results, which is similar.

currymj··on Claude 3.7 Sonnet and Claude Code
i actually don’t think nontechnical people understand the result they want of code.

have you ever seen those experiments where they asked people to draw a picture of a bicycle, from memory? people’s pictures made no mechanical sense. often people’s understanding of software is like that — even more so because it’s abstract and many parts are invisible.

learning to clearly describe what software should do is a very artificial skill that at a certain point, shades into part of software engineering.

currymj··on Claude 3.7 Sonnet and Claude Code
I think if you believe LLMs can truly generalize and will be able to replace all labor in entire industries and 10x every year, you pretty much should believe in ASI at which point having a job is the least of your problems.

if you rule out ASI, then that means progress is going to have to slow. consider that programming has been getting more and more automated continually since 1954. so put yourself in a position where what LLMs can do is a complement to what you can do. currently you still need to understand how software works in order to operate one of these things successfully.

← PreviousPage 3 of 34Next →