ivanovm

RL environment (instruction, stateful container, reward function) is the training data product being bought

No, it isn't. The private data is largely private data, created by highly-specialized, highly-paid contracted teams of experts for domains finance, swe, consulting, etc.

Reddit data is just not that interesting, that deal is worth like $60m/year. Labs spend 10x as much on computer-use RL environments.

labs invest multiple billion dollars a year each in private data, and that number is growing. internet training data is not where frontier capabilities come from, this view is outdated

I don't condone the practice, and I don't want to run afoul of anthropic or risk my reputation for it

it is certainly possible and being done all over the place. there's a black market that chinese labs use to buy frontier american llm trajectories by the millions through US intermediaries. they're not even particularly shy about it, i have been offered $0.7 per opus 4.8 call

there's also a market for chinese labs sending checkpoints to US companies to be trained on US compute and sent back

i'm surprised that so many people take chinese tech reports about how they train their models at face value tbh

The benchmarks are now the equivalents of SAT/ACT/other standardized exams for humans. They are directionally quite predictive, but with plenty of outcome variance on the margins

Yes, they did. You could also find this information easily. A company like Andon creates value by exposing interesting AI failure modes, so it makes perfect sense for them to move on to harder problems when the previous ones get saturated. I think you're just being overly cynical.

You could just look it up on their website leaderboard? The newest Claude model makes over $10k profit over a simulated year of operation, after starting with $500

I don't think this was a simple assumption. LLMs used to be much dumber! GPT-3 era LLMS were not good at grep, they were not that good at recovering from errors, and they were not good at making followup queries over multiple turns of search. Multiple breakthroughs in code generation, tool use, and reasoning had to happen on the model side to make vector-based RAG look like unnecessary complexity

Kinda funny how we went full circle with you calling me ignorant and illiterate on the basis of not using your preferred terminology, as opposed to the actual, obvious meaning of the subject

It's far more telling that you'd imagine this is a literary tool so interesting or complex that other people haven't seen it or thought of it

https://www.google.com/search?q=it%27s+not+x+its+y

Say more, I'd love to know how this is a demonstration of illiteracy

this has ai writing smell all over it. entire paragraphs that just say it's-not-this-it's-that over and over again

one underrated approach more and more people are finding success with: apple watch ultra as a primary device (optionally with a case for a more phone-like factor)

you can do most things an iphone does, but you can't doom scroll. you don't have to eject out of apple ecosystem, you get payments, 2fa, navigation, notifications. your iphone can remain as a backup that's always in sync for when you need it (e.g. traveling)

In nearly every case of RTO, there was no recorded dip in productivity associated with the move to remote work.

I am extremely skeptical of this. On the contrary there is a mountain of direct evidence that people barely work when working from home. People have been openly bragging both on the internet and in person about how they do laundry and watch netflix and mow their lawns while looking productive

All you need to do is look at the crowds in the park or lines at the grocery store on any given friday to gauge how much work is being done on wfh days

the funny part about the backlash is that the travel budgets are likely somewhat pareto distributed and under the not-personalized pricing strategy it is the little guy who is probably overpaying for plane tickets, not the corporations and the wealthy individuals

I find the very popular response of "you're just not using it right" to be big copout for LLMs, especially at the scale we see today. It's hard to think of any other major tech product where it's acceptable to shift so much blame on the user. Typically if a user doesn't find value in the product, we agree that the product is poorly designed/implemented, not that the user is bad. But AI seems somehow exempt from this sentiment

would just caching llm responses work here?

The tech industry unfortunately screwed up a basic social contract

It is well-understood in every other industry - if you want to be at a prestigious firm, make top compensation, sit in a nice office, work with top-tier coworkers and enjoy excellent perks, you must hustle hard and be unreasonably competitive every day to continue reaping those benefits.

I'm not even talking about back-breaking work - this is true for law, medicine, financial services, entertainment, sports, academia, and everything else I can think of.

After a decade+ run of cheap money and strong demand for talent, returning to broader reality may feel very unfair for many. But that doesn't make it so

I'm thinking a lot about these tools, and the biggest challenge in the space seems to be figuring out the right unit of work for AI to take on.

A signle class/function is too small to be that helpful, a whole app is too big and complex. A whole ticket is also still too big. What's inbetween? If you could divide projects into units of work of consistent complexity calibrated to AI's abilities, then you could probably get really good results.

this is very cool, have you tried DPO?

Hey guys, it's been over two months since I've been in the weeds with SCIP so I'm not going to be able to write very detailed issues, most of my experiences were with scip python and some in typescript.

1. roles incorrectly assigned to symbol occurences

2. symbols missing - this is a big one. I've seen many instances of symbols being included in "relationships" array that were not included in "symbols" array for the document, and vice versa. Plus "definition" occurrences have been inconsistent/confusing - only some symbols have those, and they don't always match where the thing is actually defined (file/position), and sometimes a definition occurrence has no counterpart in symbols array

3. the treatment of external packages have been inconsistent, they sometimes get picked up as internal definitions and sometimes not

I think SCIP is a great idea and I'd explore using it again if it got better. But I see that there are issues staying in the backlog for 6+ months which makes it seem from the outside like Sourcegraph is not prioritizing further development of scip

My, and my friends experiences with SCIP indexers built by Sourcegraph have been less than stellar. They are buggy and sparsely maintained

I've found embeddings to perform quite poorly on code because 1) user queries are not semantically similar to target code in most cases 2) often times two very concretely related pieces of code are not at all semantically similar

One of the most interesting approaches to code search I've seen recently (no affiliation) https://github.com/pyjarrett/septum

The hardest part about getting code search right imo is grabbing the right amount of surrounding context, which septum is aimed at solving on a per-file basis.

Another one I'm surprised hasn't been mentioned is stack-graphs (https://github.com/github/stack-graphs), which tries to incrementally resolve symbolic relationships across the whole codebase. It powers github's cross-file precise indexing and conceptually makes a lot of sense, though I've struggled to get the open source version to work

sorry about that, we were changing our google workspace settings recently. you can reach me at maksim@agenticlabs.com

heads up, we found the root cause yesterday's errors so should be working now

good to know, we'll give it a whirl. I tried it for like 30 mins in the anthropic web ui when it first came out and couldn't tell whether it was better right off the bat. Need to run more extended evals on that one and gemini 1.5

working on it!

we use GPT-4-turbo across the board, 3.5 doesn't handle multi-step reasoning nearly well enough to be useful here.

We'll have to figure out how to extract codebase conventions efficiently and inject them into the coding prompts. Using up all 128k tokens has given us worse results (common theme), but hopefully with newer 1m context window models this concern away