They’re both. There are finance tech bros “idea guys” whose aim is to be at the top of the company, and other tech bros who got into programming for the high salary and clout of being able to say they work at or are ex-employees of Facebook, or Google, or whatever.
While I do agree that cost per task is what customers should care about, and not cost per token. Cost per token is an objective metric. Cost to do a task can vary a lot. Different tasks, different prompts, or just pure randomness nature of models make it a bit harder to define this as an objective metric.
Unless the cost per token is prohobitedly high, people can often try the model out themselves and make a subjective judgement of how effective and efficient is it at solving tasks they usually deal with, using their setup.
Can confirm, I've got a pi 5 here that's been running 24/7 for over a year... it boots fine, but anything else makes it run so slowly because the sdcard is trashed. Its next iteration is getting a proper nvme.
I wonder, how does distillation deal with unprobed spaces in the knowledge landscape? Is a distilled model worse in some niche area that was not probed? Presumably, this is why frontier labs dont distill their own models internally to release them to the public as a servicable frontier model.
The amount of people who need and take advantage of 4x4 in a truck is probably around 5% of the user.
People mistake 4x4 for safety but really the most important part is the tires. A 2wd car with winter tires will be much safer and have more traction in the snow/ice than a 4x4 with all-seasons tires.
i let it run once to just generally find any bugs using all my quota, them once i get back to it i ran on effort /max to continue with explicit don't run workflows fix all yourself, it was ultra effective.
ultra becomes a mode where your model enhances your prompt them it is execution multiplied by all other agents, which have them going in all sorts of directions.
What about an armadillo playing a piano? There are so many potential combinations
It would say something if the pelican looked great but the armadillo looked terrible
I "grew up" using Cinema 4D. Likewise I just find that the UI of blender is for whatever reason _incredibly_ not intuitive for me. I also don't really like the fact that everything is a youtube video nowadays; I guess it's easier to monetise for the kind people who make tutorials but when I learnt 3D modelling as a teenager (in ~2005 with a desire for 3D printing!) C4D just intuitively "clicked" and they had a lot of hugely detailed, technical information with a lot of physics in it. One of my dad's friends had a license for it, and I was somehow able to just internalise its various primitives and object based workflow.
I've tried many times to get to grips with Blender but for whatever reason I've found doing the simplest things very tiring. If there's a big, illustrated PDF manual out there for a recent version, I'd love to know of a link to it...or even better yet a conversion guide from C4D (!)
"Eat healthy food and exercise" has been recommendation #1 for ages. The idea that RFK and his friends are somehow novel in this public communication is bizarre.
It's not the point of the visa, the O-1 is supposed to be for people of extraordinary ability, eg Nobel Prize winners. It's used for software engineers.
I agree. I really cringed when I viewed this platform. The potential in educational AI rests in its interactivity, quizzing and socratic dialogue. There are many great and engaging courses on Calculus taught by charismatic and interesting instructors (humans). And in this age of MOOC, all it takes is a handful of strong online accessible courses for there to be diminishing returns in trying to improve the quality of lecture content.
I want to see agents quiz me, understand my strengths and weaknesses, setup study plans, using spaced repetition to ensure I retain information, engage in dialogue such that I come to the answer through articulated reasoning. Not watch lecture content similar to what's already out there but with less soul.
1) Anthropic tokens via subscription aren't sold at a loss, they're sold at cost.
2) Subscription plans are not sold in hopes of eventually gaining a monopoly position. They act as a loss leader designed to get a foot-in-the-door and funnel companies into enterprise plans, where Anthropic can charge full API rates.
Yeah. One single project that they control. Not thousands of other people's projects breaking, I'm sure with a ton of HN outcry over releasing "this pile of vibe coded slop" upon the poor populace.
I wish Autodesk had a Creative Cloud equivalent: imagine a 50€ / month subscription that gives access to all Autodesk apps, it would have sold like hotcakes.
By remaining expensive proprietary tools, they lost the new generations of developers who learned 3D with Blender and Houdini Apprentice and who will naturally want to continue to use them once they become professionals, so their customer base will keep shrinking.
> What is the governance structure for Bun by the way? Couldn't find any documents/explanations about how it's supposed to work. I'm guessing it's essentially just "Anthropic decides what gets done and accepted" today?
And this is a valid question. You are not "defending reality" by refusing to listen to it.
In summary and based solely on my understanding:
- Jared misled people about the intentions of the migration. It's not the worst thing in the world, but it's certainly worthy of criticism.
- Jared has commented before about locking out human contributors from open source projects. Whether he was making a larger point is irrelevant as his comment stands on its own.
- Other Bun contributors, past and future, outside of those employed by Anthropic, did not and likely will not have equal access to the model Jared used for the rewrite.
Jared, working in the public Bun repository, used tooling not available to his community to experiment with a signficant migration. He dismissed all concerns and told people it's just a bit of fun, and that it shouldn't be taken seriously. Most of the controversy would have been avoided were the experiment done in private.
None of this is a big scandal but questions about the project are entirely justified.
Have you actually used Opus 4.8 in Claude Code? It takes way too long to do any practical task on higher thinking levels due to over-engineering. And I am not the only one complaining. Lots of people downgrade to Opus 4.6 exactly for this reason.
Opus 4.8 training works well for agentic work. Not for code harness.
A local AI is not about cost. In fact you will likely pay more for it than with most providers. Just look up the advantages of having access to a technology like this that can be self hosted
Expecting that governments won't respond to software which is tailored to prevent them from investigating crimes (e.g. Signal's primary marketing is all about preventing government access to chats) was so utterly naive that only privileged techbros could make it up.
Other people understand that picking a fight with someone will result in a fight.
Fuck me, you HN "entrepreneurs" are really delusional. And then you wonder out loud how come us, the poors/normal people hate on AI and on the tech class as a whole.
Back to your point, you are aware that many ot today's wages are far from covering one's basic expenses, aren't you?
Good move given some experienced issues and compaction across the 5.6 range is closer to 5.4 than 5.5, i.e solid and reliable.
Will say that 5.6-Sol is a minor bump in my benchmarks in most areas vs 5.5 but a severe regression in a few specific task focused on rearranging trees, addressing merge conflicts, etc. where the model to accomplish the task does not properly adhere to prompts in a way GPT-5 originally managed, not retaining parts of history in the way prompted despite specific instructions not to as that made the final completion easier…
I am of the conservative and cautious opinion that no model should be able to run destructive tasks at all, I have seen every model do things that make me concerned enough to maintain that opinion and know my evals can’t catch everything. But for 5.6-Sol specifically, I’d caution everyone to reevaluate how you run the model, maybe take a few more precautions you tend to forgo.
It is extremely capable as a reviewer and for extensive tasks, though for the later, the safety net I feel is required to be comfortable limits the utility.
I find the 4-bit QAT with MTP to be entirely usable speed on both my boxes (Strix Halo and a desktop with two V620 GPUs, which are slightly faster than the Strix Halo).
It’s also like smartphones. In the early years, every year was a huge jump. I still remember marvelling at my iPhone 4’s detailed display, and video calling for the first time.
Now? I don’t even know or care about what the latest iPhones have, I’ll get a new one when mine breaks.
This is the correct answer. `ic(foo(123))` can be written as `print(f'{foo(123)=}')` without depending on yet another third-party library which is not pulling its weight.
Seeing all the people up in arms that there was an automated rewrite of Bun to Rust has been very revealing and disappointing. Very silly behavior and denial over the inevitable.
For whatever reason prefill (on my DGX Spark) is way way faster with the Gemma models than Qwen 3.6 models of similar size. On vLLM anyways. Likely just deeply tuned code contributed to vLLM by Google?
vLLM gives me 7000+ tok/sec with Gemma 4's MoE model. That's... pleasant.
The shit and exploitation don’t build themselves.