ltononro

Worth noting/understanding, for those that are not familiar with it, that these 33k tokens are not a single 33k batch of tokens, they persist and increment in every request.. so if you have 10 requests that is 330k tokens, 100 requests 3.3M requests, 3.3M that could be ~5x less if used another harness. These 33k tokens are mostly because the harness is completely bloated with mcp, skils, plugin, loop, basic alignment and all sort of explanations needed in the system prompt for it to work the way anthropic wants it to... imo this is sub-optimal and if the model needs that much orientation, it says a lot about its intelligence too.. Good models should be as harness agnostic as possible

they are playing with being fast, not good lol

Grok: "I can answer any math question in less thant 500ms" User: "what is 2319321x232" Grok: "3201521321" User: "this is wrong" Grok: "but it was less than 500ms""

Using LLMs to judge this could be very hackable, shouldnt it be? Is that the best practice we expect? Would be very interested in see some ablations about failure modes where the LLMs tried to hack it somehow. Or failures from the llmaaj

I wonder about the motivation to switch teams. What has Google done wrong? Was he tired? He could retire, open his own lab, raise capital. So many opportunities, why go to OpenAI? Folks talking about the amount of money paid, wasnt he the guy that was acqhired for billions? would OAI pay billions (basically to google) to get him?

lol so weird to put that into perspective macbook ~8b airpods ~18b iphones ~60b

I know they are massive, but AI seems something much more important than airpods

The headline is a little bit misleading and sensationalist imo Most of the loss comes from the non-profit->for-profit switch This is not operational loss, one-time stuff to make the company for-profit and compete with Anthropic in apples-to-apples comparison, am I right?

I wonder how one gets data for training foundational models like this. Any idea?

This is one of the things I am most afraid of. Governments can break the progress of AI and this could be a bubble burster?

Well but comparing with sonnet 4.6 instead of opus 4.6,.7 or .8 doesnt make a real point I mean, pay 200 USD/month (if you have that cash, or your company has it), might not justify using local at all (unless you have some reason to suspect about data leakage)

Good depends a lot. If you are in the token maxxing hype you will probably find these models very bad comparing to SOTA, unfortunately.

The good news might be: opensource models are now good (enough) for day2day usage. But is it really? I feel that companies will always naturally strive for the best and use the SOTA (as long it is not too expensive).

I see OSS models being a good backbone for companies in the future that have validated workflows and could use those for privacy or to spare costs.

IDK, might have gone a little bit off-topic here.

Building a network is also very important. It is a long run and also a habit to keep

Well I wouldn't email anyone. LinkedIn, X, WhatsApp and even reaching out through friends seems much more effective But the key is to get to know anyone that you think it is minimally (or surprisingly) reachable and try to get the best out of them. And always be reachable up to your limit. Help people whenever you can to reach their goals, you'd never know :)

https://huggingface.co/yuxinlu1/gemma-4-12B-coder-fable5-com... That might be the closer we can get to it lol

So you don't really trust the data policy (non-retention) of the big companies like Anthropic/OpenAI + regulations in EU. This is very interesting. I myself have been blindly trusting these organizations with my data and still not sure if I am trading code/trajectories for productivity.

Another POV is that most of the code written in most of my codebases were generated by Codex/Claude, so they would be "stealing data from themselves" in a sense.

I've been working with Transformers/LLM training in 2018-2021 and then now, more recently again. Things are far different. I think they would be more interested in the "how" you got your code to be satisfactory with your guidance than the actual code generated. But mostly I personally trust that they are not really using my trajectories for that (unless I explicitly allow it in the configs)

It's hard right now. I've been interviewing some folks for a while.

I am really interested how someone can make a 100x of themselves when working. This often means "how are you using claude code", "what is your coding setup", "how do you decide if your agent is going in the right direction or not w/o looking directly at the massive amount of code generated". This is not technical, but it measures how much we would have to teach someone that comes from a non-industry background. Remember that companies have money, universities dont pay for a 200 USD Codex+Claude+Grok+Cursor per employee.

That said, I also had to ask the same stuff to folks from big techs that dont actually are software engineers. How are data scientist leveraging LLMs to automate experiments? Are they using it for something other than data viz? Are they using it for /loop exploration? how trustworthy are the agents right now for this?

Having someone with this grasp nowadays might be very very valuable IMO

Well that is like 3 USD/hour if you run it on a rented gpu

What kind of coding do you do? Do you keep track of frontier models to vibe check the differences and re-evaluate constantly or are you ok with having a nerfed model forever? (not being judmental, just really wanto to know your framework here)

controls could be easier :D