randbyte
randbyte on Hacker News
If you look at the open weights official checkpoints they are either fp8 or even 4 (kimi). None is using bf16 for inference.
It’s just a different format. Providing liquidity in the hope that more is generated from the economy. Except it’s doing more harm than good other than in true crisis. Other times it’s just creating a bubble, in national debt and in AI.
Every FED governor is appointed by POTUS.
Way more people complain about this too, if you pay attention to all the US inflation and national debt anxieties. It’s clearly unsustainable and driving interest rates up steadily.
“All major labs use SFT” <> it’s effectively for reasoning. And SFT <> distilling.
lol looks like you know about model trainings from water cooler chats mostly.
These two slimy companies do belong to each other. Perfect match made in hell.
lol. Have you? What kind of distill you are talking about? It’s only effective in the surface, small scale model or output preference. Not useful for long reasoning chain.
lol. Yeah just use “CCP” and “security” as bogeyman keywords you’d absolve all the crimes from closed labs and win arguments automatically
Except you don’t have the choice to “pay 10% extra just to buy from Amazon”. Amazon force sellers to set not-lower price everywhere else.
On the contrary. There are plenty of choices on iOS but no diversity.
There is no browser engine diversity on iOS. Reskinned Safari doesn’t count.
Who is vals.ai and why they keep submitting eye-catching claims. A few weeks ago they said fable 5.1 solved some obscure cipher and now opus 5.5 found room temperature semiconductor candidates. Meanwhile they seem to be in the business of making benchmarks.
Are they a promoter / influencer for Anthropic?
Half a million loc to do the basics feels like highly inefficient. Pi 1.0 also turn on the “fullscreen” tui by default… it’s becoming more and more like opencode.
This is naive. With extreme prevalence of vibe coding it’s a matter of time before someone turns their local app into an mcp proxy.
I don’t know. They are self hosted so I am not comparing token cost.
ds 4.1 is lightning fast though. Also much better in image recognition.
Certainly no “anyone” and who happens to be worse than no one.
Just like how Anthropic and OpenAI is already doing?
4.1 flash is very fast and capable. Token efficiency is not great so it fill up context window much faster compared to similarly capable models.
glm 5.3 flash is a tad slower but a bit more capable and way more token efficient.
Source: self hosted tested on rented GB200 node at 8bit.
When Chinese tech report is tech report and US tech report is bible scripture.
Blatantly lack of principles.
Long history of open source projects already have answers on how to monetize a free and complex open source product.
- Low development cost: collaborative efforts from open source contributors, innovative model training and serving for llm (Chinese models costs a fraction to train and their local chip design and manufacturing are catching up, plus cheap electricity)
- monetizing by selling hosted services, while leaving the core product free to tinker with / self host. China’s gdp is 2/3 of the US and it’s already a huge market for AI - which OAI and A\ don’t enter.
- for (the US) market that they can’t enter, let the US cloud providers to do free marketing / advocacy for them. Gaining share of mind. It costs them nothing.
When people say “US government” they usually mean the federal government…
Bibi tried many times. Every president before Trump understood the risks and refused the wild plan. Only Trump fell for it.
China’s factory outputs are in fact increasing. So is their export. You wouldn’t see this if they are energy constrained. China’s problems are weak consumption and deflation. It started way before this year and has very little to do with Iran.
You think too much. Trump got duped into a gamble and didn’t win as he thought. If the hope is to hurt China it’s unlikely to be working very well. Their energy sources and oil sources are very diversified.
In your example, service providers may be allowed to collect IP address or cookies or referer which are either critical to how they provide service (IP and cookies) or part of the established norm (browser sending referer).
That doesn’t mean they are allowed to go beyond and scan user’s environment variables
> a model that runs on my own machine will never have the capacity of a model that runs in a datacenter.
I don’t think so. A local run model only needs to serve one or a few people. It seems possible to run a DeepSeek v4 model at full capacity on a server costing 200k usd. Very expensive but not impossible.
Factor in hardware and software improvements over time, and the fact that most people may just need to run a smaller and quantized model, it should take a pc at 10k usd scale.
> Releasing open weights seems to be an effort to slow down competition.
Why? I don’t get it. Open weight models enabled a lot more foundational model trainings. I believe R1 was benefited from llama.
The proliferation of alternative, open weight models in turn put heat on the leading labs and forced them to make better models, or squeeze more out of slowing model capabilities through innovation on harness.
And eventually we the common people benefit, exactly from these competitions
Instead we got home solar system become very affordable over the past decade.
And driving out US manufacturers isn’t even the main goal for China. They know their huge risk on reliance on petroleum and was doing everything they can to mitigate that. Building out a huge solar manufacturing base is their answer. Now they are reducing petro imports YoY.
I never say these pieces are the equivalent of a cloud platform. What I said is that with them, it is entirely possible for someone to build their own.
Why someone wants to do that is not my concern. But not being able to is.
If you see this as people just want free stuffs you are missing the point. There is an entire different level of danger to allow a few entities to completely own a day to day technology.
Imagine personal computers and open source operating systems completely don’t exist. What you can do on a pc today, you have to always do it on a thin client always connected to a main frame run by a selected few. Everything you do is recorded and subjected to at least 30 days of retention and inspection.
Imagine every car is completely not customer serviceable and have to be connected to one of the three manufacturers in order to operate. Everywhere you go is recorded. The manufacturer may decided the place you want to go is inappropriate for you to go.