casualscience

There was never any plausible explanation for why this wouldn’t happen.

What a nice post hoc revision of history. Distillation is still an active area of research, that you can distill models as easily as you can it genuinely interesting and absolutely not something that was taken for granted even 12 months ago.

Even 6 months ago this idea that 'using model outputs as training examples' was listed as the reason that all models would fail in the near future due to some spooky circular training catastrophe.

Don't pretend like this was so obvious.

Honestly, 60-70% of this feeling is just inflation. If things were still affordable, the doom and gloom would be much less palpable. Inflation will get under control, but we still have other problems with the direction of social cohesion.

brain was grown in house and likely deserves as much or more credit than open ai for the current llm boom

Search, yt, etc.. are the real revenue sources for Google. But they are serving [3.2 quadrillion tokens per month](https://blog.google/innovation-and-ai/sundar-pichai-io-2026/...) and have ramped up a ton in 2026, yet profits continue to expand.

Comparing to ad revenue from a company like meta, the story that Gemini tokens are a strong cash drag on Google just doesn't add up. It seems at worst they are losing like 50 cents/1M tokens (including r&d spend, data centers, etc..), and very possible they are actually profitable per token.

Which is much better than anthropic and openai.

How are the Google numbers calculated? I've seen their net income increasing a lot as they've rolled out Gemini. This suggests that Gemini tokens are actually profitable, or at least not extremely unprofitable.

Yet this site suggests that tokens are very unprofitable

If you actually have time to read all your code, understand it, and are willing to be bottlenecked by human understanding, then yes, you are living in a different world.

In my world, that is far too slow, and you will be seen as a low performer who just can't keep up with the tech.

Not my experience, most people are mercenaries, most people make mistakes, most people need their code and architecture reviewed. Obviously working with a person and working with an AI are not identical, but most of the broad responsibilities are the same.

And it's much more like senior IC than it is manager.

While this isn't a unique perspective, I think it's wild more people don't understand this. What happened is everyone is being "promoted" to staff+ level engineer and they're realizing the realities of that situation.

The funny part is that these are the same people who are upset that these folks up the food chain "do nothing".

A lot of people focus on the whole dynamic of corleone family and their position in American society in the Godfather. But seriously did you see what vito was eating in the restaurant?? Makes no sense

Kojima always has some paranormal activity in his games. And he does like to subvert expectations about what is technology and what is paranormal

Not much, just accurately predicted the next 30 years exactly

Are these meta engineers that were let go? The one thing you learn more than anything else as a Meta engineer is how to sell your work and how amazing it is

Idk, do you think it's sensitive for the employer to train an AI with it and then put that AI on Instagram for everyone to use and ask for employee SSNs?

They absolutely do, wtf are you talking about.

Also people use their work accounts and laptops to read their w2 and other sensitive info.

How does it not apply anonymity? All you see is some vote came in with a particular public key. That's it. Thats more private than we have now

Even in small groups, being respected and considered valuable is important? I'm not sure what you mean here.

I take your point, and I too get triggered when people invoke mate selection and dopamine. I could be with you in being skeptical about that specific angle... but absolutely if you look at lawless or less institutionalized cultures, there is a trend towards appearing strong/tough and hiding any weaknesses

Victim blaming cancer patients as cope so you can convince yourself "it won't happen to me" is a disgusting trend

Okay, average person uses a special key picked up from the DMV one time that allows them to login to vote.com and cast their vote. This is a totally normal experience and understandable by anyone who has done online banking.

Google's profits have been going up while 'giving away gemini for free', so I don't think they're 'being irrational', they're unit economics apparently work.

Hmm, whats stopping you from running claude code on a separate machine you can ssh into? I don't understand that point at all, I do that all the time.

Using a claude code instance through a phone app is certainly not something that is easy to do, so if there's like a phone app that makes that easy, I can see that being a big differentiator.

I see, so there's actually an additional for loop here, which is `sleep(n); check_all_conversations()`, that is not something claude code does for sure.

As far as the 'soul' file, claude does have claude.md and skills.md files that it can edit with config changes.

One thing I'm curious about is whether there was significant innovation around tools for interacting with websites/apps. From their wiki, they call out like 10 apps (whatsapp, teams, etc...) that openclaw can integrate with, so IDK if it just made interacting with those apps easier? Having agents use websites is notoriously a shitty experience right now.

what is the memory architecture, doesn't this already exist in claude code?

Can someone explain what value openclaw provides over like claude code? It seems like it's literally just a repackaged claude code (i.e. a for loop around claude) with a model selector (and I guess a few builtin 'tools' for web browsing?)

People succeed in spite of these systems. They have resources, tremendous network advantages, and the people at the very top crust of engineers are indeed quite good at their job.

Crapping something out as quickly as possible and leaving somebody else to deal with the fallout of a bad data model and violent on-call isn't something to be rewarded IMO.

Sadly you've described precisely the optimal engineering strategy for promotion at my FAANG

Also lets 50% of errors through

Honestly, I feel like I have to know more and more these days, as the ais have unlocked significantly more domains that I can impact. Everyone is contributing to every part of the stack in the tech world all of a sudden, and "I am not an expert on that piece of the system" no longer is a reasonable position.

This is in tech now, were the first adopters, but soon it will come to other fields.

To your broader question

Something that I think many students, indeed many people, struggle with is the question "why should I know anything?"

You should know things because these AIs are wrong all the time, because if you want any control in your life you need to be able to make an educated guess at what is true and what isn't.

As to how to teach students. I think we're in an age of experimentation here. I like the idea of letting students use all tools available for the job. But I also agree that if you do give exams and hw, you better make them hand written/oral only.

Overall, I think education needs to focus more on building portfolios for students, and focus less giving them grades.

Yeah I hear you. My understanding is that on youtube you can make ~2k per 1M views with the default ads. I'm hoping that I can be funded by some combination of that and something like patreon/membership/merch. But we will see, it's something I've wanted to do for years and I am getting too old to put off longer.

These are impressive metrics, are you able to make a living off of your 10M views?

I'm planning to leave my job this year and focus on content, mostly have been considering YouTube, but if blogging can work too, might consider that as well

If you work at a megacorp right now, you know whats happening isn't people deciding to use less libraries. It's developers being measured by their lines of code, and the more AI you use the more lines of code and 'features' you can ship.

However, the quality of this code is fucking terrible, no one is reading what they push deeply, and these models don't have enough 'sense' to make really robust and effective test suites. Even if they did, a comprehensive test suite is not the solution to poorly designed code, it's a band aid -- and an expensive one at scale.

Most likely we will see some disasters happening in the next few years due to this mode of software development, and only then will people understand to use these agents as tools and not replacements.

...Or maybe we'll get AGI and it will fix/maintain the trash going out there today.

Also worked with Soumith. The man is a legend, moves mountains and completely changed the course of my career because he liked something I wrote. No arrogance, no politics, just an extremely down to earth and chill guy who elevates everyone around him.

Hope him the best!