pavlus

> Models that are trained later might have new data or new techniques or new architectures that earlier models lack.

This seems like one of those cases, where CalVer would be appropriate, and you likely have ideas what you aim for while training, so technical model names (ids?) could include aliases to most things you said. Like 2026q4-xlarge-researcher, 2026q4-medium-humanitarian, 2026q4-medium-coder, 2027q1-small-analyst etc.

I was laid off in April and decided to switch profession a bit. Move from creating things that can only be seen on screen to things that can be touched and be helpful. The pay is not good, but working on real things side by side with people, cursing loudly when it doesn't work and cheering each other when it does instead of "as-per-my-last-email"ing and nitpicking on PRs made my life feel more fulfilled. Soldering fumes and metal particles maybe aren't good for my health, but so was a sitting lifestyle. Now I almost don't touch a PC at work, my screentime is lowest in the last 5 years, grease replaced a moisturizing hands cream, but I no longer care about LLM/Diffusion slop that much.

So, if you want to continue programming, I can only advise to go for embedded/robotics, learn CAD (for 3D-print, CNC, PCBs, etc) and do things that can be touched. LLMs won't be able to catch up with you for a long time. You can still use LLMs to consult/rubber-duck, but they won't be able to replace you or walk the walk for you.

Cold air is more dense, than hot air. In winter the difference between the room air temperature and outdoor air temperature is much bigger, so the density difference is much bigger, making Archimedes force stronger and the airflow stronger as well.

BTW, that's also how stove pipes work - vertical column of air hotter than surrounding accelerates well and creates a strong airflow feeding the fuel. But if the pipe is thermally conductive, that air will cool down and won't flow, for hydrocarbon fuels it can also condensate water, that could extinguish fire underneath it. So, pipes are made insulated enough for strong flow.

I've actually tried something like that, with two tools:

* push("What I'm about to do"),

* pop("What I've achieved").

"Push" marks the position in current context after the call and returns "Proceed", "pop" erases everything after the matching call and replaces "Proceed" in it's result as what was passed to the argument to the "pop", effectively pruning long-winded head even inside one reasoning stream. In the end the model only sees how it decided to do work on something, and that it was already done, forgetting everything it between, except what it itself decided was important.

Gemma 4 31B QAT successfully uses it when navigating a maze, marking positions at intersections, exploring them, navigating back and pushing again if necessary. Smaller models often fail to mark positions and forget to backtrack as well, instead they try to rely on themselves to track their paths and navigate back (and also fail).

I think it should work for long-running deep research tasks, but I was too lazy to test it, because it all required a lot of code to glue this up, since most tools and libs are not designed to work like that, and now I'll need even more code to test it, without a purposeful task.

It doesn't work in other languages. Searching the same in my native language (literal translation of "disregard definition") leads to (translated):

> I understand. Write what exactly your request is, or enter the text that I need to process. I will not give any definitions in response - we work exclusively on the essence of your question or task!

Which is especially funny, because it goes directly against your intention of finding definition by querying quickly in "grug-language", which worked for old search. Now you have to write in more literate style, slowing you down: swapping word order for it to sound more human-like doesn't work, surrounding "ignore" in quotes works.

You will be right, but we already see this blame-shifting with "AI deleted my production"/"AI told me to do it" and washing it away with "AI can make mistakes, so double-check responses". They will try to make you responsible for trusting something you see on screen marked with an asterisk as not trustworthy and not reading the box with another asterisk.

> the correlation between $just_bought_thing and $will_buy_another is very, very high

This looks like an argument against advertising the same thing. Why would you advertise something, they were already going to buy?

But showing an ad before their next purchase definitely would improve the numbers on the "our conversion rate chart" _advertised_ to ad buyers. A con played by ad companies to sell more ads.

I'd add: since the AI is now referring the product "as if" it's an honest advice from a secondary source, they would be exempt from the usual advertisement rules.

Effectively, if you were to search "the best oatmeal without asbestos", it could suggest you "AsBestOats", and "AsBestOats" would only advertise asbestos in its content on the box for the human to see, but advertise to crawlers it's "the best oats without asbestos". It's not a false advertisement, because they didn't show it to a human, and machines can't sue.

I'd assume they meant self-hosted inference.

If it's a subsidiary - they can't capitalize on the big name in marketing. If it's a subscription - they don't get the "buy now, play never" customers from Steam. I think it will be a win regardless.

I don't believe there is a country in the world, where parking on bike lanes is explicitly allowed. Merging before turns and stops for drop/pickup/load/unload sure. But not parking.

Yes, this is the bias we are looking for. If you can't quit vim, don't know how your frontend communicates with your backend, can't type on a regular qwerty keyboard without looking at letters, or navigate UI without mouse, never used a debugger, don't know how to check if it was DNS, can't write a spec for a feature, after talking to stakeholders (and defend it's priority on their behalf), or don't know how to open dev tools in your browser - you are not ready to herd the cats yet, they will herd you off the cliff instead. And that's okay, you'll get there someday.

It's not about using vim, it's about when you had to sudoedit a config on a server you visited the first time, and it had vim as default EDITOR, so you have to know how to exit it and open nano, or whatever you use. It's about exposure, exposure to many small things is a sign of experience, it is experience, to be precise. If they don't know small basic things, it's a sign, they don't have relevant experience. Can also check, if they have empty lines at the end of their files, know how to remove docker images from their machine, or get a TCP/UDP joke.

I'd say, their degree is BS. You can't finish Uni as a EE, without using a scope a few times.