f311a

I hope they will put more resources into fixing issues. I remember there were multiple attempts to add to ClickBench, but each time new bugs were found.

It should not be X times slower than SQLite.

https://github.com/ClickHouse/ClickBench/issues/336

Europe doesn’t even have cheap electricity

What is Antares? Can't find anything related to it except for the 1B model, which does not seem to be capable of autoresearch.

UPD: It's his personal bot.

Mine sometimes contains funny takes.

Something like: "I have been looking for a QA middle role with 2 years of experience for six months. Here is the stuff I won't be doing: relocating to a different city, commuting more than 30 minutes, ...". Most of the comments say, "Good luck. You will definitely find a job". Only a few mention that it's time to update your worldview.

Never seen such posts, my feed contains different kind of larping.

It’s better to compare this to computer science, especially, theoretical one. LLMs stop people from exploring new languages, architectures and so on.

You can use ClickHouse cloud to host it on AWS.

But given the majority of use-cases of CH, AWS can be quite expensive.

They got like 10 new sponsors, you can see it on their page. They are one of the top OSS projects in terms of revenue.

I can't recall any other OSS project that makes so much money and has nothing to do with infra/cloud.

They were making around $1M a year back when they complained about the revenue. A lot of Linux distros can only dream about such an amount of money.

A lot of open source that we used to use as dependencies is trivial. Sometimes you only need 10% of the library. No one is talking about kernel.

Also, changing license does not prevent you from forking the source code before the change. There are no new libraries with kernel level effort that use uncommon license.

Well, people don't play like computers, so this kind of practice can be useless.

Because we use regular CH in production. I'm pretty familiar with it and I can always run large queries on a production server with 1TB of RAM.

Regular CH also support external data sources, so I can read 500GB of JSON from S3 and group by it on production server very fast and in memory.

I've been using clickhouse-local for quite some time, instead of DuckDB. There is also chDB.

After using pandas for 10 years, I favor SQL now, for some reason.

Is the google infra stable enough right now? At the start of the year, the flash model was unusable for a whole month via gemini CLI. They could not fix it for a whole month and I was a paid customer.

I feel the same, there are still lots of people who are interested in learning internals, but they are hard to reach. Even such people don't use Google anymore, so they won't find your content.

Why not? There is $700k reward for the next iteration of this benchmark https://www.kaggle.com/competitions/arc-prize-2026-arc-agi-2...

I would not call this overfitting, it's finetuning for specific task where you have a benchmark.

The whole point of his model is to optimize for a very specific benchmark.

BUT, he does not use labels when training, so the model does not know the answers.

Robotics cost way more to produce and support, unlike software

Unless it's a critical CVE, they're worth very little now for your career.

Same for a lot of side projects or small contributions to OSS, people are still chasing them and wasting everyone's time for nothing. They wouldn't even work on them before AI.

There are more than 10 crates for arenas with different tradeoffs, I'm well aware of them. They are still very limited compared to C/Zig.

You can’t allocate collections without nightly or without reimplementing them in the library. Every implementation uses it’s own set of trade offs to provide safety in unsafe implementation.

Stabilized soon, really? They did not stabilize it after 10 years and were thinking about different approach. I thought it’s dead.

He won’t reply, he’s busy promoting himself and his peojects with AI.

Unfortunately, Rust is not a good choice for this kind of tricks. This is where Zig shines. In Rust, you can’t even use proper arenas, which can help a ton with allocations.

Cloudflare started to pick Zig recently, for projects, that have memory constraints.

This does not account for anti fingerprinting. Also, are there any new tricks that are not in fingerprintjs?

Prior to LLMs, I've rarely seen such headings when reading blogs, but now they are everywhere.

From the article:

> A kiosk that kept dying

> “It’s my device”

> Reality television

> The grind

> The relief pitcher

This is one of the giveaways for me.

How is the general knowledge of Qwen 3.6? Do you need to explain things outside of algorithms to it? Since the size is so small, I guess you need more explanations to it. General knowledge helps with coding when your don't specify a lot of details and ask for big changes.

Why is it so small, but expensive?

Open Router

Input /M $0.45

Output /M $3.20

Cache read /M $0.05

Throughput 27 tps

It would be a very nice model at 200-300 tps and if it was dirt cheap. What's the limiting factor of optimizing speed and price for inference providers?

Opencode said they are working on matching the old prices using their own inference.

Right now, they give 4100 credits for Luna and 63 000 for Deepseek on their prepaid plan (both are 2x)

It's just model size and heavy RL, sometimes they overfit on specific tasks. RL can get you very far, prior models did not have such a focus on RL for agentic setups.

Look at deepseek, they improved it just by doing a lot of RL and you can see it from how it behaves. You provide very little information about a task, but since they are trained on similar tasks, they come up with a lot of assumptions and details on their own, because they were trained with such an info during RL.

The only problem I have with LLMs in Go is that they also make a lot of concurrency mistakes, in the same way as people. It's easy to fix though, just by asking to double check the code