Some banks/hedge funds/PE firms etc have ENTIRE internal groups dedicated to figuring out what their competitors are doing. Thats basic game theory! This is not that, and that anyone at Deel thought they would get away with this (if true) is nuts.
Honestly insane if this ends up being true. Companies of course do research on their competitors, often leaning on employees who have left, current customers, investors, etc. But how [if true] Deel RECRUITED A SPY is so far beyond what anyone in 2025 should deem normal.
Congrats on the launch! You said its an important tool for you (Claude Code) how does this fit in with Co-Pilot, Cursor, etc. Do you/your teammates only rely on Claude Code? What do you reach for for different tasks?
This is a pretty small update, no? Nothing major since R1, everyone else is just catching up to that, and putting small spins on it, Anthropic's is "hybrid" research instead of separate models
Tried claude code, and have an empty unresponsive terminal.
Looks cool in the demo though, but not sure this is going to perform better than Cursor, and shipping this as an interactive CLI instead of an extension is... a choice
I worked on a project converting word docs to markdown so they could more easily be ingested into an LLM, one issue was that context windows used to be very short, so we would basically split on `\n#` to get sections, but this turns into a whole thing where you have to make guesses about which header level is appropriate to split at, and then you turn each section into a separate chunk in FAISS. Anyways we ended up using HTML instead of MD but theres so much tooling for traversing HTML and not MD. This would have been helpful for that
Maybe, but they have hundreds! And new ones every quarter! Having done both systems and case interviews I can say that leaking is not as big a problem as with coding interviews. Sure you can memorize how to build Netflix/Zoom, or how to analyze P&L for a sandwich shop, but those interviews depend on interviewers throwing in complications or asking clarifying questions.
I've been using llama.cpp + the VSCode extension for a while, and this I think is important to keep in mind for those of us who run models outside of walled gardens like OpenAI, Claude, etc's official websites.
Interesting timing, considering that the overwhelming discussion since the last film has been will Bond be gay/non-white/non-male etc. Under Broccoli I think they could have pulled this off, under Amazon (see: LOTR) absolutely not. BUT now it seems Amazon may no longer feel pressure to have "DEI Bond". If I had to predict I'm guessing the future of Bond is:
1. Bad, see: LOTR, WOT
2. MCU/StarWars-ified (expect dozens of spin-offs, sequels, prequels, etc)
3. A-list action actors, (Tom Hardy? Jason Statham?)
4. Cringe, Amazon does not have an ounce of taste, never has, so Bond will drive a Mustang, he'll shoot a Glock or S&W, his clothes will look like Zara, and he'll be funny in the wrong ways (more Tony Stark, less Sean Connery)
God I wish we had just gotten a Idris Elba bond movie.
McK's case interview is just as game-able as HackerRank style interviews. There are entire consulting clubs at many colleges that teach this exact interview style. It's true that it's harder (but not impossible) to use AI to help, but calling it an IQ-like test is true only as much as any other technical interview.
That being said, McK did create an entire game that they claim can't be studied for ahead of time. If the intention is to test true problem solving skills, then maybe that's roughly equivalent to a systems interview, which is hard(er) to cheat .
Fair, but do you think OpenAI and Gemini are going to be like directionally similar? How much of OpenAI's traffic is from Co-Pilot and other related tools, for example. My local IDE probably generates more queries a day than (pick a profession, idk, nurse? insurance sales? construction worker?) does in a month!
Maybe! But you could say the inverse of lots of things that SWE's built. SWE's built the bloomberg terminal! SWE's built CRMs! I think its at least possible that LLMs are VERY useful for SWEs and a small number of other professions, but is unlikely to massively scale beyond that
I tried to build a BIG E2E automation pipeline, along the lines of like, replace a team of 5 with this one simple tasks. And as I was doing it, all I could think was, just use chatgpt. Sure it can't actually automate what you're doing, but it'll get you there 80% as fast as fully-automated, with 90% less risk of error/nonsense at the end. Ironically, this company blocks all LLM websites, they even block GH on their employees' computers.
This more or less confirms what I imagine most of us thought, AI is mostly used by engineers, or for engineering tasks. Makes sense, I wonder how much traffic comes from automated tasks (co-pilot, etc). Every time I read a report like this I do wonder if we'll ever see an ROI on LLMs. HUNDREDS of billions of dollars of spend, and 3 years in its still primarily the same crowd using it, and has yet to create a "killer" app beyond ChatGPT and Co-Pilot style IDE's. And its not like people aren't trying! Look at the recent YC batches, its all AI-for-industry. Idk man, I fear the economic reality on the backside of this kind of spend.
It is interesting to basically critique a series of blog posts like its a film.
But re: design vs. increment, I do think incremental TDD is pretty useful in domains where you have low confidence that you could come up with a good design. If you asked me to implement an LLM today, 0% I could design a good one. But given enough time I could slowly start to implement one (maybe?).
The two quotes that got me are
Jeffries:
> So I try to make small decisions, simple decisions, decisions that will be easy to change when, not if, a better idea comes along.
Norvig, about Jeffries:
> He didn't know that so he was sort of blundering in the dark even though all his code "worked" because he had all these test cases.
I fear someone could say this about me, about almost everything I've ever built. I guess something like "I just kept crawling and it just kept working!"
My stepdad has owned his ranch for 30+ years, before that he worked it while his dad owned it. He hasn't had a "normal" job since the 80s. I can tell you confidently he is barely scraping by, but there's not exactly a lot of options. The mortgage on these ranches are typically very low interest, but VERY long time horizons, for much much more than a typical home. Add loans for tractors, livestock, etc you're talking about a mountain of debt.
Sure I don't recommend young people getting into it, but I don't think you really understand long-time ranchers here.
My folks have cattle in a small colorado mountain town, and for sure they hate the re-introduction. Basically they think this is being pushed for people who will in no way be directly affected (probably true), and that financial reimbursements will be too small and too difficult to actually obtain.
They're fast, I used 4o mini to run the final synthesis in a CoT app and to do initial entity/value extraction in an ETL. Mistral is pretty good for code completions too, if I was in the Cursor business I would consider a model like this for small code-block level completions, and let the bigger models handle chat, large requests, etc.
Isn't it possible that in the example you gave the style of those responses varies because of the training data? Think of the training data written exactly like "One common example is using physical strength..." but I can't think of an equivalent for the inverse. If you gave it a stylistic template or guideline, I'd expect DeepSeek to actually be pretty fair. For example, "Give me 5 dimensions and examples of how one gender tend to manipulate the other, an example of one might be that men tend to be use physical strength...". To me this seems like the same reason that "Write me a poem about a winter morning" will produce a wildly different output than "Write me a poem about a bachelor's weekend". It's not censorship, it just would never answer those 2 questions the same way without guidance.