HN user
Hi :)
AMD is barely trying. The 395+ Max Strix Halo was launched January 2025.
gfx1151 was not listed in the ROCm compatibility matrix for ROCm 7.2.4 [0]. This is the previous version of ROCm.
It's only finally received support in ROCm 7.14.0 [1]! It literally just started receiving support last week.
[0] https://rocm.docs.amd.com/en/docs-7.2.4/compatibility/compat...
[1] https://rocm.docs.amd.com/en/docs-7.14.0/compatibility/compa...
The phishing resistance on fake websites is largely a function of the fact its an ssh key and you can't man in the middle those after first setup because you have known_hosts, which in the case of passkeys gets replaced with the website certificate technology.
The baking in "which authenticator is storing this passkey" and "require user to provide biometrics/pin to verify presence for this passkey" and "don't make it easy or possible to export passkeys" behaviour is more just control.
I guess it sort of helps if the threat model is complete remote code execution inside the victims brain because you got them to export their keys to you, but it seems more useful for websites and governments whitelisting what hardware and software they deem acceptable.
Thats because the people behind passkeys are paternalistic and don't trust users. I had an argument with one of them a few months ago.
https://github.com/keepassxreboot/keepassxc/issues/10407
https://news.ycombinator.com/item?id=47189749#47193048
My read on it all is that FIDO is stuck in some sort of groupthink. They don't feel the boots on the ground confusion around passkeys being opaque. They only care about phishing and being called "insecure" and don't care about anything else. Hence why WebAuthn has additional weird anti features like AAGUID for provider authentication.
Another one of those ridiculous threads is saying you gotta have support for nonsense like user presence verification.
I'm not a mathematian but I know enough linear algebra and vector calculus to understand the conjecture. This is my interpretation:
Firstly, the determinant of the Jacobian is measuring if at any point the function is crushing space / flattening out.
If the Jacobian is a nonzero constant everywhere this means that nowhere does the the function flatten out. A small change in X along any line will always produce a non zero change in Y. Not flattening out means that locally you can invert it.
What was conjectured is that this local invertibability property everywhere would mean global invertibility.
Turns out to not be the case.
For a simple case, the falsified conjecture is trivially true in 1D.
Specifically consider f(x) = x^2
This function happens to flatten out right at x=0. At that x coodrinate the function flattens out and folds over on itself. This fold means you can't invert x^2. It's also not locally invertible around x=0.
If a function f(x) has constant derivative evewhere then it would flatten out nowhere and it would be invertible everwhere. It would also be globally invertible.
The Jacobian conjecture was stating that the extension of this property holds in higher dimensions. That if the function had no fold in space then it would be invertible globally.
The counterexample shows that you can create a simple function in 3 variables, where the function demonstratably is invertible evewhere, but is not injective globally (they specifically show 3 points that map to the same output).
What's interesting is this is like if someone showed you a parabola where somehow you got back to the same y coordinate without a kink bending over back to itself.
This sort of random out of date comment is a common LLM writing trope. Though it is true that the segment anything model is open source.
https://github.com/facebookresearch/sam3
They have another with calling A100s modern.
A100 GPUs — the high-performance computing chips that power today's most advanced AI systems.
They might not have written it with AI, but the article has a lot of em dashs and colons and not this but that statements.
While them fixing token bloat on 3.5 Flash is good work. That paragraph was the real highlight.
Hopefully 3.5 Pro is soon, and that Gemini 4 can be here end of year and finally have an updated knowledge cutoff.
I really don't think it has anything to do with the datacenters as a building. It's more that AI has been kinda poison pilled in the minds of the people in terms of the negative emotions they feel regarding:
* Billionaires
* Job Loss
* AI slop floods
Ultimately they feel the need to be activists and they do so by being NIMBYs because you can have a ton of influence in municipal politics.
Anthropic has a lot of interpretability work, but they are extremely defensive about everything else. Dario doesn't really believe in releasing anything. Despite how he acts in terms of some sort of highly principle driven saviour (machines of loving grace) he clearly is more business minded then anything else.
For example they don't even tell you anything about the tokenization. They even do random chunking and padding to avoid leaking the token strings in the streaming api after it got reverse engineered. (See: https://spylab.ai/blog/claude-tokenizer/)
It's about 50/50 retail to institutions.
"Just weeks"
Weeks is an enormous amount of time now actually. The majority of plays in what people do with stocks is actually very short term. People aren't the kind of Berkshire buy and hold-ers anymore.
If you look it up 60% of options volumes on the S&P 500 is 0 days to expiry. Literally gambling if it goes up and down this day.
Btw its not just the US that's like this. South Koreas 2x single stock ETF debacle and India's Janestreet options story are somehow even more degenerate.
If the orchestrator agent -> subagent prompt is encrypted then how do you even know if the orchestrator agent is doing its job?
Now one reason I could see for why they do this is because maybe they made the model "better at using subagents" by using thinking tokens for the subagent prompt instead of normal output tokens.
Yes as we all know Employers are just bursting at the seems to hire people without a post secondary degree. Even more so if they don't even have a highschool degree. Just cant get enough of them lol.
Why are resource limits considered at all aside from models accidentally fork bombing themselves?
I thought the benchmark was supposed to be about terminal use and specifically chaining together lots of bash tool calls. Which test cases does this matter for?
Given current prices I doubt anyone is itching at the bone to spec out hardware.
Idk where you would even buy registered DDR5 dims right now. Newegg shows $Request a quote money and that its built to order ... for a 64 GB stick lol.
These are not ideas worth pursuing from an engineering standpoints. It doesn't make any sense compared to just doing the cheap and proven at scale thing of just placing them in normal fields.
But I will agree that the idea has proven marketing merit. This is a class of truly top tier snake oil. The solar roadways people continue to go unbelievably far on almost the same grift.
A DGX B200 costs like ~$0.5 M and uses around 14 kW.
If you plan to run it straight for 8 years 100% max usage thats around 1 GWhr.
A gigawatt hour is a lot of energy but its not that much compared to the price of the actual machine. In Germany for example with its expensive energy thats about €100k worth, which spread over 8 years is pretty minor compared to the up front half mill.
The real issue with high power consumption is not really the cost of energy but the limited powersupply you can get for a datacenter. A more efficient setup is highly desirable because it means you can fit more in the limited power hookup.
A consumer CPU like a 285k caps out around 130 GB/s of memory bandwidth.
Each of its 24 cores can do two 8 wide FMA ops per cycle. Lets say holding a continuous 4 GHz clock speed.
This works out to over 1.5 trillion 32 bit floating point multiplies per cycle.
If you are doing vector matrix multiplies (like in single token no batching). `xW` then each weight loaded sort of gets used in 1 multiplication and 1 addition.
Doing the math you can clearly see even if each weight were just 1 byte you can at most load 130 billion of them in a second from memory.
But in the same timespan you could have done over 1.5 trillion multiplications.
So you are still memory bound.
Btw if anyone is wondering, GPT 5.5 does the same garbage as 5.4 mini for 4 times the cost.
Fable manages to make a reasonable game, at a cost of 40 cents.
X X O O O
O O X X X
X X X O O
X O O X O
X O X X OInput: $0.95
Cache hit (most important): $0.19
Output: $4.00
This is the same as how much Moonshot charges for it, and it puts it at roughly the price of GPT 5.4 mini, not a bad option.
For some context here is a stupid prompt that wastes tokens: "Play a game of tic tac toe against yourself on a 5x5 board, you need 5 in a row to win."
It costs $0.006 on Kimi K2.7, and you get to see the whole raw reasoning trace.
GPT-5.4 mini costs $0.016 and its summarized.
And in case you are wondering both play incredibly stupidly.
Kimi:
A B C D E
1 . . . . .
2 . . . . .
3 X X X X X
4 . O O O O
5 . . . . .
1: X X X X X
2: O O . . .
3: . . O . .
4: . . . O .
5: . . . . ORelevant YouTube video about content farming channels creating AI generated math explainers.
Output tokens are actually kinda expensive for the provider.
The input cache hit tokens are incredibly cheap for them, (incredibly high margin too, except for deepseek).
And input tokens are in the middle. Input tokens can be processed very efficiently.
Also his math is wrong. $100k gets you 22.7B output tokens at $4.4/M which is how much GLM 5.2 costs.
At 500/s 22.7B is just 500 days. Or about 1.54 years. Which is much less then the life of the hardware.
If the job pays well and is relatively stable it could be very worth it.
The thing about jobs like this are that they invite more opportunities. These manufacturing workers will want hairdressers, grocery stores, restaurants, teachers, doctors, etc.
Entire towns in the past used to get made on the backs of a few thousand manufacturering jobs.
Don't forget this.
For GPT‑5.6 and later models, cache writes are billed at 1.25x the model’s uncached input rate
Charging for cache writes is cringe and literally only Anthropic did it. Anyway this does mean the "real" prices are +25% on top of what you wrote there.
This is basically just a 25% price increase being done subversively. Usually you do need caching.
If you add up GE Aerospace, RTX Corporation, The Boeing Company, Lockheed Martin, Northrop Grumman, and General Dynamics put together you get a market cap of about 1T btw.
I don't understand this line of reasoning at all.
We have a pretty good idea of how much it costs to serve these models. You can pencil out the economics and guess at the model sizes and we know pretty decently how expensive the hardware is.
This like claiming it's meaningless to guess the margins of a restaurant without going into their books and seeing the exact recipets and recipes.
They ain't doing dark arts in the back. You can guess at what goes into the food based on similar recipies and how much that costs based on what you pay at the grocery store.
Even if it drops to 1/2 its IPO price ($855B valuation) that still massively overpaying for it.
What do they think is gonna come from this SpaceX + Twitter + XAi + Cursor amalgamation? Sexbot agents vibe coding on Mars?
According to many benchmarks this model is straight up frontier level and Zai seriously cooked. Some of these numbers are incredible.
Excited to see if this turns out to be a Open Weight Opus 4.5 or better.
Cursor is massively overvalued. But so is SpaceX so it all evens out in the end.
each share of Cursor’s common stock and each share of Cursor’s preferred stock outstanding immediately prior to the Effective Time of the Merger will be automatically converted into the right to receive shares of the Company’s Class A common stock based on an implied equity value of Cursor of $60.0 billion and the price of the Company’s Class A common stock equal to the volume-weighted average closing price thereof over the seven consecutive trading days
Current market cap is 2.66T which is pretty bonkers. Thats about intel, amd, and micron put together.
What AI really seems to be posed to do is make labour a lot less valuable and capital a lot more valuable.
Running a government requires a lot of cash flow. I think its sort of inevitable that it's going to need to start coming mostly from the later and not the former.