fwlr

You never hear anyone talking about “imposter syndrome” anymore, do you?

    An AI agent, 600k tokens into a feature that it just authored, understands the feature as well or better than I do.

No, the driving through Chinatown question is a question for self-driving cars. It is not a question for LLMs. There is some other question for LLMs, and the author is using the ancient technique of analogy to get you to think about that question.

Very concerning - I’ve tried n+m, n*m, n^m, and even m^n, but I can’t seem to reproduce the popular “10x” claim.

Best of luck!

    If enough capital has been installed before learning removes the wedge, the economy lands in the high-capital state,

Post a picture of one of these with the “X” key conspicuously removed and you’d probably get a repost from Sam

Vyvanse is dexamphetamine. I suppose you can say “amphetamines” instead, I imagine the subtleties of the terminology differ quite a bit in different places.

Strattera can help (I mentioned bupropion, another SDNRI). Of course if you try stimulants and don’t respond well, it’s totally fine to just use e.g. Strattera. What I’m advocating against is e.g. “try Strattera first, if it seems to help, don’t bother with stimulants”. (Some places, cultures, or medical systems do have surprisingly strong biases against stimulants!)

    This could be due to ADD, I am still getting tested. Granted, that's a diagnosis, not a root cause.

The treatment for ADD is one of two medications, methylphenidate or dexamphetamine. You can try other things in addition to these, but not instead of these, and you should try both - there really is just no substitute.

(In some places, bupropion can be prescribed as an antidepressant. It has effects that also help with executive dysfunction, so you may find it to be more effective than serotonin-based antidepressants.)

Having trouble finding the motivation to grade that fifth essay in a row that starts with “Certainly, here’s a possible introduction to your essay”? Just ask Claude to grade it instead!

Naturally it depends on how well Cloudflare built this implementation. In the abstract, though, a sufficiently accurate prediction system should expect to recover the causal structure of the phenomenon it is attempting to predict, and thus an imitator hoping to defeat the predictor should expect to contain the same causal structure (i.e. physical simulation of a human hand to arbitrary level of detail).

Near the end of their methods:

    For each likely AI-generated target, we search human controls from the same repository and same programming language, controlling for project-level practices … We also apply size matching: a control must be within 20% lines of code of the AI target

    Comparing the lines of code (LOC) between AI-generated and human-written files, we observed no statistically significant differences

    For mechanical work–things like bumping versions, fixing failing tests, checking documentation for inconsistencies–the clankers have proven their worth (even before LLMs, anybody remember dependabot?)!

    The way to fix that, of course, is automated processes and mechanical guardrails. Automatic enforcement of style, automatic testing and linting and formatting–same as ever. 

    We can use the whale fall tokens to build that machinery.

Granted, but we’re talking about agents, which fundamentally presumes the obsolescence of concepts like “end user” and “good software”.

What do you think a tool call is? What is this small but crucial difference between a `read_file` tool call and `scp`?

If your contention is just that this should upload files one by one instead of all at once, what you want is for providers to facilitate the illusion of privacy.

Well said. I hope one day it becomes possible for users who choose to install and run said software to also be able to remember this.

The jitter you add has to specifically be “jitter that mimics human cursor movement”, which is extraordinarily non-trivial to synthesise.

If you decide that your entire home directory is the project, as the OP did by setting the repo_path to ~/, then, well… I mean, if you ask me, I don’t recommend it, but it’s your computer and your free will.

It’s not really paranoia that makes me expect this will happen. It’s more like, well, the model weights and the files I want the model to work on have to be inside the same GPU for the agent to actually work, right? So step 1 has to be either “they send me a server rack of GPUs”, or “I send them the files I want the model to work on”. I’m not sure I could reasonably expect anything except this to happen.

Is the Grok CLI a 2 terabyte install? Did Elon dropship you an 8U rack of B200s?

No?

Well the model weights, the GPUs, and the context obviously all have to be in the same place, so “sending your project to them” is literally the only thing that could possibly happen, unless you think agents work by fucking magic.

This is the biggest case of PEBKAC in history, maybe ever.

This is the kind of confusion that Charles Babbage could not rightly comprehend, except at those politicians at least had the excuse that computers had only been invented five minutes prior.

I get it, I want to agree, I really do like the “this is a new tool in the toolkit of the professional software craftsperson” argument…

…but consider: the Q-tip. “Don’t use it to clean your ears”, but for most people that’s all they want to do with it, and empirical observation indicates that this dynamic results in either “using Q-tips irresponsibly” or “not using Q-tips”, with “uses Q-tips properly” being a small-to-vanishing proportion of the whole.

Please don’t discourage this kind of public thoughtful introspection.

Was hoping they’d link this to the Overview Effect and they did. Very plausible explanation, I like it.

The author considers the possibility that this is an outwardly-polite form of “you should be able to solve this yourself”. Another, grimmer possibility is that it’s an inwardly-polite form of “I should have been able to solve this myself”.

That is, the author asked for

    the thing 30 years had taught him that a search engine couldn't

    I have forgotten that thing

Presumably, some initial information was fed into the start of this reporting process. Multiple stages of this process had near-total incomprehension of the information yet performed full ingestion and reconstitution of it anyway, leading to this terminally-confused output.

I found this post very refreshing! I’m sure it would have been very tempting to one-up the “PR-speak” of the Bun post. Likewise, it would have been very tempting to include the same set of facts that reflect negatively on Jarred, while studiously concealing one’s own opinion (eg “I heard people called him a stinky manager. I am not saying that, other people are, but I’m not”). I appreciated that it was just … genuine.

I think we more or less agree, with the caveat that I think social engineering attacks are far more worrisome and threatening than SQL injection. The gold standard solution to sql injection (prepared/parameterized queries) is guaranteed effective, and does not impede the efficacy of SQL. The gold standard solution for social engineering attacks (role-based access control with minimum rights) is only almost guaranteed, as the attack could be made against the management or admin who ultimately holds the keys to full rights, and most certainly does impede the efficacy of the humans operating under it.

    “Prompt injection attacks have become, to agentic AI, what SQL injections were to web applications: a systematic, category-wide vulnerability class that requires the same systematic strategies and defenses.”

Isn’t prompt injection far more fatal to LLMs than SQL injection is to SQL databases?

Like, the problem of SQL injection was that user input was forming part of the instruction string given to the SQL engine, and so malicious user input could include various SQL grammar terminals to end the current SQL command, followed by complete SQL commands of their own, and the engine would simply execute both commands. The fix was prepared statements: fixed/static/pre-compiled instruction strings, that can only ever perform fixed/static/pre-defined logic, and that logic can then be (more) safely applied to arbitrary user-input data.

The analogous mitigation for agents is to have fixed behaviors they can perform, such as “read repo 1” “read repo 2”, etc., and the user input is used as data to select which of these fixed behaviors to execute. But we already have this technology - it’s called a menu. The value of LLMs is specifically and intrinsically predicated on being more than a menu, while the value of SQL does not depend on being more than “pre-set logic operating on arbitrary data” - user input being part of the instruction string to SQL was incidental, for developer convenience.

I do see the appeal, it’s easy to imagine that workflow working, and working well - but it’s hard to how it avoids this fate: https://youtu.be/QEzhxP-pdos

Thank you for mentioning zmx. Reading through their repo, it’s like someone read my mind, took my instinctive, almost pre-verbal, half-formed/quarter-baked preference for `screen` over `tmux`, beautifully and precisely formalized it into an actually lucid opinion, and then instantiated it as an actually real piece of software. I’ve used it for an hour and it already feels like second nature; I somehow have muscle memory for features that I didn’t even know it had until I instinctively used them and they were right where I expected them to be. Seriously, thank you.