hbbio

Actually, the "narrative" crashed Nvidia for no reason.

Not only DeepSeek uses a lot of Nvidia hardware for the training.

But even more so, by releasing an open weight frontier model, people around the world need more Nvidia chips than ever for inference.

Someone should try to use a local LLM (maybe a 4b) to merge the diffs in case of conflict beyond straightforward cases...

Not energy efficient but should work surprisingly well without CRDT, OT, or anything else.

That's why in practice you need more than this simple loop!

Pretty much WIP, but I am experimenting with simple sequence-based workflows that are designed to frequently reset the conversation [2]

This goes well with Microsoft paper "LLMs Get Lost In Multi-Turn Conversation " that was published Friday [1].

- [1]: https://arxiv.org/abs/2505.06120

- [2]: https://github.com/hbbio/nanoagent/blob/main/src/workflow.ts

Yes, agent loops are simple, except, as the article says, a bit of "pump and circumstance"!

If anyone is interested, I tried to put together a minimal library (no dependency) for TypeScript: https://github.com/hbbio/nanoagent

Spoke to some people working at big banks: They say trust, or rather lack thereof, _between_ banks is a key issue. Blockchains solve that.

Good point.

It seems the common approach now is to make a v2/v3/etc. of your protocol and let your own users migrate. Previous versions will still run forever, but your frontend can push migration paths.

I would say most current blockchains are pretty bad at privacy. Look at what onchain sleuths do.

Rather, blockchains enable anyone to provably check the state of the database. Your fintech app logs you out. No proof, and frustrating customer service hell for days/weeks/months.

Social recovery for smart accounts covers your use case

Somehow yes, and it's quite sad.

The blockchain technology itself is a marvel and the safety/security properties of properly decentralized systems are unmatched. The whole financial world would be better written using blockchain/smart contracts, same goes for voting/governance, etc.

But these scammers. So many of them... Easy to smell though, but the problem we need to fix is these scammers get most of the online traction/attention, taking it from truer builders.

Glad to see this trending on HN! As I was born in France, I had the pleasure to read it. Just thought the multiple translations available on the site seem like a good corpus to see if frontier models could improve the translation.

Below is o3 take on "Le Chat", given as prompt the French original and all existing translations. I am not an expert in poetry and maybe not versed enough in English poetry specifically, but it looks suboptimal: It changed the structure, some verses seem overlong and I don't find the original beauty in "barbed claw's art".

https://fleursdumal.org/poem/132

Le Chat

Come, lovely cat, upon my ardent breast;

Sheathe in your velvet paw the barbed claw’s art,

And let me drown in eyes where, coalesced,

Cold agate gleams within a molten heart.

--

While idle fingers roam and fondly chart

Your supple head and sinuous arched spine,

My hand grows drunk on thrills that softly start

Across your vibrant body’s living line.

--

Then, in my mind, my woman’s gaze is mine:

Like yours, dear beast, it pierces—deep, serene.

From head to foot a perilous airs combine;

A subtle scent swims round her dusk‑brown sheen.

Now it is, thanks for suggesting. Qwen3 4b seems to be the best default model for usual steps.

https://github.com/hbbio/nanoagent/pull/1

Thanks, /nothink works!

So, Qwen3 1.7b is about the same speed just slightly worse than Gemma3 4b which is pretty impressive.

Qwen3 4b passes all 200 tests and is much faster than Mistral Small 3.1 24b or Gemma3 27b.

Thanks!

Turns out just is not the word here. My benchmark is made using conversations, where there is a SystemMessage and some structured content in a UserMessage.

But Qwen3 seems to ignore /no_think when appended to the SystemMessage. I can try to add it to the structured content but that will be a bit weird. Would have been better to have a "think" parameter like temperature.

Feel free to add a PR :)

What is the parameter?

I also have a benchmark that I'm using for my nanoagent[1] controllers.

Qwen3 is impressive in some aspects but it thinks too much!

Qwen3-0.6b is showing even better performance than Llama 3.2 3b... but it is 6x slower.

The results are similar to Gemma3 4b, but the latter is 5x faster on Apple M3 hardware. So maybe, the utility is to run better models in cases where memory is the limiting factor, such as Nvidia GPUs?

[1] github.com/hbbio/nanoagent

From the article: "the impacts of generating a 100-word email. They found that just one email requires .14 kilowatt-hours worth of electricity, or enough to power 14 LED lights for an hour"

Seems completely off the charts. A 70b model on my M3 Max laptop does it for 0.001 kWh... 140x times less that stated in the article. Let's say the OpenAI Nvidia clusters are less energy efficient than my Macbook... but not even sure about that.

Indeed, that's not clear... was trying to teach no rodents!

Crows soak all their food in our (sorry, their) swimming pool! Sadly, that includes rodents too. I wonder if it is because of the chlorine that could potentially "clean" the food, or if they want to wash out sand, etc.

Been careful not to yell or approach abruptly and they definitely learn to recognize our faces, since we can get pretty close to them now.

Not explicitly written in the summary but this seems to be mostly for Python right?

Guessing from the reference to TACO, LiveCodeBench and the mention of a runtime sandbox for Python.

Congrats! I like Zig a lot even though I never implemented a full project with it.

FWIW we had to build a related lib in TypeScript: https://github.com/okcontract/graph as part of the runtime of https://github.com/okcontract/cells

Since Svelte 5 was announced, we began using our own library for state management as we could not fully understand the semantics and were afraid of potential changes. If you're curious, here it is: https://github.com/okcontract/cells

It is compatible with both Svelte 4/5 and more importantly pure TS libraries.

"My first prompt created a CoT so long that it catastrophically forgot the task"

Many humans would do that

Thanks!

If the M3 can run 24/7 without overheating it's a great deal to run agents. Especially considering that it should run only using 350W... so roughly $50/mo in electricity costs.

How would you compare the tok/sec between this setup and the M3 Max?

Too bad the evil one won

At first I thought the title was about _software libraries_.

And I wasn't so sure the answer to that could be definitive!

Same, I have a Mac but every week I check the progress of linux support for the Snapdragon X1E!

I am running Sonoma (and iOS 17) with no plan to update. I don't want AI built into the OS. Very happy with ollama and our own code.

In fact what I would like the most is to strip the OS to the bare minimum: Remove Photos and whatnot from the OS (hey it's an app, I should be able to install it from the AppStore if I want) and the hundreds of daemons that run to do who knows what.

I tell myself I'm a minority here... but everyday I stumble upon all types of users who want the same.

> I can remove that ingredient but I would loose a lot of sales.

Or: I can remove that ingredient but it goes against my principle of not accepting constraints.

> Some people died. Was it my fault, knowing how it was being used? Did I do anything illegal? Unethical? Immoral?

https://www.imdb.com/title/tt15398776/

Of course not, the tweet here is _the_ original source for that news

A drug trafficker sells drugs

A developer builds a platform like eBay but without censorship that can be used by the drug trafficker

It's not the same thing