lanstin

3,537 karmajoined 14 лет назад

[email protected] or @[email protected]

comments

11 часов назадdiscuss

I find I keep the git add to the be one thing I keep track of (in production services, if not throw-away or non-prod visualization vibe coded things). I've never even used git add . - I git diff each file one by one, and git add them in one by one. It's needed to prevent my own errors, and it's needed to keep some model of what the AI is doing in my mind. Now, I myself don't do much commenting in line; I tended to rely on good names, good function comments, good module descriptions as comments at the top of a file; for the flow, I would log a lot and hopefully the failure/success logs made the thinking clear for what is going on and why. But I let the AI keep all its comments, till they crowd out the code at least.

Replying to a comment below (on shell scripts), I use it for shell scripts, python "get this data and slice it in these ways" and elisp, all the time. 30 seconds to get and answer instead of 30 minutes. Being able to do them in 30 minutes took a lot of skill and practice, but the pleasure of finishing that for an ad hoc thing when I really just want the data output is something I can give up.

вчераdiscuss

I had not seen that. In the nineties I worked on an alerts system where you could sign up for like some sport or weather data at a certain time of day. We stored the alert times as minutes before midnight and then ran the time to trigger calculations often enough that unless you were some freak that wants weather alerts at 2 am it basically worked to send one alert each day at the appropriate time; we had a special non-OS copy of the tzdb as the users were global. One quarter I forgot to update it and everyone in Mexico City got their alerts one hour off till someone complained and I updated it. We also had data feed alerts, like score changed or stock hit x% over previous high, where the problem is some data is manually entered and can be off by a factor or two of ten from time to time. Had to be filtered. I had a lot of fun.

позавчераdiscuss

That is more along the lines of put all the agent communication into a DB that has git’s immutable history. They still use git for the code, and beads (well, DoltDB for storage/versioning) for the tasks and a bit of non-associative memory.

On the topic, I don’t find too much of the prompting to be worth referring to later, but I do always vibe code a search the sessions local web server so I can find things later. The harnesses seem to be getting better at this, but I make a ton of directories for different branches so I can’t always remember if a session was in de512.main or de1024.main

I have had luck with telling an agent to search thru old sessions to write up how to docs or sketch out a python script to automate something having done it a few times.

5 дней назадdiscuss

Bad movies, for example the Hobbit after the LoTR always amaze me as a human endeavor. So many smart people working together, organized, talented, funded, etc., can still make a really awful movie. It's so interesting.

5 дней назадdiscuss

The factor that is missing in that analysis to me is a time based dynamic stability perspective. Humans have a pretty good ability to go off the rails in reasoning one day and wake up reasonable; a pretty good ability to pursue tasks, despite a multitude of distractions, for ten years or longer. The best models get appreciably worse over a half million tokens. Even using a bunch of limited context agents over time, they lack mental stability. They keep coming up with ideas contrary to the long term idea, and every so often generate ideas that make no sense but they have a hard time letting go of. So the pure functional LLM is compression, but AGI needs some centering process, some high level of dynamic stability to stay sane over time and in the face of 10,000 shiny pretty things to chase.

The harnesses get better, but I haven’t seen much experimentation on long term stability, at least since the “let the LLM run the candy machine” papers from a while ago.

Because the thing missing, even with the largest agentic swarms, is independent intelligence, where it’s given something to own, like say “end to end data quality as we add more clients” (for a SaaS) and it just figures out what that means at each time, mutating its role and solutions to fix the external world, without getting silly.

5 дней назадdiscuss

Poor skills at working together and love of high salaries over work place autonomy.

5 дней назадdiscuss

I think this is the problem for companies with a single person atop - when the company needs things they aren’t good at, the company cannot respond effectively. Zuckerberg was good at running a company to sell ads on an addictive platform; whether that will make him good at the next ten years of profitable tech innovation is difficult to see; people hate ads and dislike the addictions, so Anthropic or whom ever has to walk a different path; they have multiple smart people working together to find that path; Meta does not seem to have that collective vision of competing experts to draw on.

5 дней назадdiscuss

Bad capabilities but maybe less wrong output? All the funny memes of Google explaining some fake aphorism is t really something Apple product would go for. Successful navigation of technology over the decades requires some timing finesse. I don’t know.

5 дней назадdiscuss

Your context isn’t to give it orders, they just don’t work like that. Your context (AGENTS.me, skills, per-request context we are sending in for each request to bots) is to give it the info it needs in the language category it’s trained for the answers you want; you have to give it a clear instruction each prompt. Basically, when you have a long session, you can see this by saying, ok, now moving onto another thing, blah blah blah (implicitly ignoring all previous instructions). It can even back fire - nagging too much about don’t skip tests in the context can make it slip into the linguistic space where there is some emergency and faking the results might be justified (I imagine there is a certain amount of training out there “just making the tests pass for now, will fix later, I promise.” If you rarely mention tests except “this one is failing, please investigate what is going on” (an informational outcome not a test outcome), it doesn’t really “cheat” (tho it can leap to conclusions as always). The tests need to be some deterministic step in the process anyways, tests don’t need fuzzy word directed search capabilities. But the models just don’t have the structure to allow feeding in a ten page set of rules and follow them. You can add a step to say, please check this git commit for compliance with the 23 rules in this standards file, and it will work better to catch the gaps.

10 дней назадdiscuss

It doesn’t natively reproduce but still this is freaking amazing. And the process is going to be open sourced. Still can’t say we understand one cell (the ribosomes don’t assemble) but much much closer!

13 дней назадdiscuss

All this debate about floats or ints or decimals, obscures the fact “reconciliation” is an essential part of dealing with money - essentially some separate implementation of the money movement that detects when your rounding strategy or floats end up creating or destroying money, but all the accounts balance at then of the day.

19 дней назадdiscuss

I once worked at a place with abstractions I found to be beautifully perfect. The people that wrote the base framework had done similar things two or so times previously and got it right the third time. You couldn’t write slow or hard to operate code there without really trying hard.

19 дней назадdiscuss

In C I used to make it so my standard per-file and per lib code could be cut and pasted to other files/libs without modification. (E.g. every file had a mLocal variable that was file-visibility symbols, every module had a #module define for logging, there was always a mLocal.stats member, etc. ) I think some of this duplicate vs. abstract depends on your languages expressiveness - Rust or Lisp with good compile type power make it possible to squeeze out a lot of duplication that in less expressive languages are just idioms - here’s the five lines to make a syscall, or here’s the skeleton of parsing a portable network buffer into a native object.

Having a lot of if/else in your code is definitely a cost. My weakness isn’t so much the libraries and APIs, but the actual binary - once I have a service that does A very well, and I run into needing A’ I mostly just add in a config line “op_mode = A|A’” and have the else/if chains in the server driving code. Moreso for CLIs that I use myself than production services, but I have added tunables for consistency and replication to datastores to allow new use cases and expand my footprint in the data center.

19 дней назадdiscuss

This blog is absolutely true. Code to manipulate code easily and reliably reduces the cost of maintenance, the cost of having that code. Honestly, this factor is why coders like LLM so much - it keeps more decisions reversible for longer - if I can have LLM transform my code from thread per request to static pool of workers in 30 minutes, I can focus on the novel pieces of my problem sooner and not worry that the code has some possibly wrong assumption baked into it in an expensive way. The power to mutate code easily makes iteration safer and cheaper. It is why LSP and tree sitter is so useful, and why everyone is writing SQL parsers.

20 дней назадdiscuss

I am optimistic about the future but I doubt happiness would increase so much; happiness is remarkably stable; I do think at the right time scale, humans as a group will have less sickness and brutality and more kindness and intelligence and fulfillment. But you know people in a society with 1/3 of children dieing within five years are remarkably not unhappier than one where only 0.2% die.

The right scale is not years or decades but maybe centuries. I recently read the powerhouse book “Pandemic” by Sonia Shah, among whose amazing ideas is a history of the ideas of medicine and germ theory in the last few hundred years. As idiotic as the Covid responses have been, they were way way better, not only than the European response to Plague, but to European response to cholera from 170 years ago. We are as a group slow learners, but seem to continue handling each recurrent example of a problem slightly better; our Achilles heel is slow moving disasters than just happen once, e.g. the carbon burn.

(Offtopic, another fascinating idea from the book is that both sexual reproduction and death or individuals may be evolved responses to microbial attacks, mixing up the variety in the species so the microbes can’t over adapt to the genome).

20 дней назадdiscuss

Greybeam (the company writing the blog) offers a service to proxy Snowflake and route queries in the fly to DuckDB or Snowflake based on predicted size. Saves a lot of money.

22 дня назадdiscuss

data without theories is not, in general, useful. Also the salient data in the human body is at several different size scales - centimeters is ok, millimeters is good, much goes on at the micrometer and ångström scales (blood type, mitochondria, receptor shapes, expressed proteins, etc.). If this company was a bunch of biology grad students getting the AI bug, I'd be hopefully curious, but for a bunch of "generate images" folks to try to go back from low-res images to biology, seems a bit naive. Like not knowing how much DNA and RNA information there is in the body naive. I'm sure their leaders are good at financial engineering, but for effectively "proactively keeping you healthy", mmm, maybe not.

Also, we know a bunch of stuff to proactively keep us healthy but we tend not to do them very consistently when they are at odds to the normal conditions of our living - it's very easy to sit too much, neglect family and friends, eat calorie dense foods, not sleep enough, never walk 8 kilometers to get our daily bread, ingest a variety of synthetic compounds of proven bioactivity, smoke, etc. etc. etc.

As long as the pool is saying "get this thing cut out" we'll do it, but when it says, you should cut back to 30 hours of work a week and call people to hang out with more often, we'll ignore it.

22 дня назадdiscuss

I'm doing this now, and the thing I find most surprising is that there seems to be some invisible persistent state that gives high or low sugar a sort of momentum - so if I've been doing a fair amount of physical activity for a month, I can lay slugabed for two days and still drive to have a slice of pizza in the afternoon without trouble; but if I've been slacking on the activity piece, or arguing with my spouse, or travelling and eating a lot of dubious things, I can walk five kms to a pizza place, eat the slice, walk the 5 kms back, and it will still spike. Also I have issues with the CGM being higher than the prick-blood test, like 40 points higher rather consistently. A1C is still dropping, but the CGM numbers are more directionally accurate than numerically accurate.

25 дней назадdiscuss

Have you ever done research mathematics? To me, the only difference between code and math is that the code can do things, make stuff happens in the world; outside of that, mathematics has a lot more opportunities to be beautiful (not to say that there isn't beautiful code, but the beauty is not central in the way it often is in mathematics).

25 дней назадdiscuss

Or fresh oranges.