HNHacker News
TopNewBestAskShowJobs

Alpha3031

315 karma · joined September 24, 2019

submissionscomments
Alpha3031··on “Next-token predictor” is the wrong mental model for LLMs
Diffusion LMs denoise a canvas which I personally find more interesting.

I don't really disagree that human cognition is essentially a predictive task though, as I understand it, predictive coding and related theories based on the Bayesian brain hypothesis are fairly popular these days (though maybe not clearly dominant over alterative models? IDK I'm not a neuroscientist). I imagine most people would draft a few tokens before refining them like MTP or diffusion though, if we do decide to use LMs as an analogy to human cognition.

Alpha3031··on Go grandmaster Shin defeats AI KataGo with a two-stone handicap
Probably Fox would be the server with the most players but at DDK basically any would work and I think getting exposure to different playing styles would be worthwhile.
Alpha3031··on Go grandmaster Shin defeats AI KataGo with a two-stone handicap
Maybe not to the extent you mentioned earlier (I definitely agree that it's not "focused" on this) but there is definitely asymmetric playout training in the training data (see comments on lightvector/KataGo issues #39 and #162 mentioning it for example), that is presumably how they got the tweak for playoutDoublingAdvantage (i.e. actually having a few thousand of the millions upon millions of training games be games where playouts have been doubled for advantage).
Alpha3031··on Gemini 3.8 Flash and 3.8 Flash Cyber
AU user also, just checked AI studio since that seemed like the best bet and both 3.8 and 3.7 show up (and can be used for chat in playground, though IDK what the limits for that are). Chat in gemini.google.com is also 3.6 for me but I'm on free tier lol so I don't exactly expect it to show up any time soon. I think there's also another free API beyond the AI studio one (which is 20 RPD free according to docs so not really useful) but I forgot where it was (Google cloud maybe?) and what the limits for that were.
Alpha3031··on Gemini 3.8 Flash and 3.8 Flash Cyber
The search model is probably flash-lite based on what they give to users who aren't signed in.
Alpha3031··on Nvidia's AI advantage is moving beyond the GPU
Personally, I don't think it's (paid) PR per se. Call me cynical but I'm fairly sure TechCrunch has always been mostly lightly laundered press kits. This is, of course, something that is endemic across most trade or trade-adjacent media.
Alpha3031··on Hy4 preview
IIRC Nvidia claims to release enough data that it should be possible to fully reproduce Nemotron, so even if it's not as good as the current best models, GPT 5.1 or Opus 4.1.was still useful right? I guess it depends on what you wanted to do with them.
Alpha3031··on 9th Circuit sides with states in Kalshi gambling fight
Well, one should also remember that they thought protecting the opulent minority was so important that an entire house ought to represent them. Secure the permeant interests of the country and all that.

Not that the other house isn't also tilted that way, but it seems to be that they, and the appointments approved by them, are functioning exactly as was intended (as shown by the supreme court suddenly caring about executive overreach when it came to money). Who knows whether they would care about the social issues affecting the poors nowadays, but they certainly didn't care to let the people be in charge without a veto from that opulent minority.

Alpha3031··on GLM-5.3-Flash
Normal MoE is switch-weights-per-token so you would se substantial slowdowns that way. Apple did a More that switches weights per prompt (instruction-following pruning, https://arxiv.org/abs/2501.02086) but you have to design the model that way which I don't think the have.
Alpha3031··on Xiaomi: New CPU matches Apple cores single threaded, much faster multithreaded
Xiaomi are implementing ARM cores on an ARM ISA so any architectural defects would also affect other ARM processors.
Alpha3031··on Devices with GrapheneOS support should be available in 2027
I am not a bank so my being aware of things is not going to change much.
Alpha3031··on Devices with GrapheneOS support should be available in 2027
It would be funny if it actually was intentionally selected, but for DES-like instead of Dual EC reasons, and they're just too embarrassed by Dual EC to admit it.
Alpha3031··on Devices with GrapheneOS support should be available in 2027
Some banks are app only and/or build MFA functionality into their apps.
Alpha3031··on Sticky wage norms and the real wage cost of unexpected inflation
"Idiosyncratic" doesn't mean "we expect it to me less accessible", it means "we expect it to behave in its own special way so just taking a random one of them doesn't tell us anything about the average experience".

Given the number of goods and services there are in any given economy, you can probably find at least one good of each of the following categories:

A) that was very inaccessible in the 70s, and accessible now,

B) that was very accessible in the 70s and still very accessible now,

C) that was very inaccessible in the 70s and still very inaccessible, and

D) that was very accessible in the 70s and very inaccessible now (in addition to the one that you already picked)

None of those four tell us anything about whether "things people usually purchase" or "things the economy generally produces" have gotten more or less accessible, because different goods have gotten more or less expensive at different rates. You can easily point towards increasing inequality, which shows that the median American is getting a smaller share of income, and the divergence between CPI and GDPIPD, which serves to show that what people purchase on average (CPI) are getting more expensive faster than what the economy produces on average (IPD). (by about 30% on average according to FRED)

Insisting on intentionally using less useful evidence to support your claims is like making a strawman of your own argument, when clearly a stronger argument exists.

[FRED]: https://fredblog.stlouisfed.org/2023/03/the-differences-amon...

Alpha3031··on Sticky wage norms and the real wage cost of unexpected inflation
I am taking your reply as stating a lack of intention to provide any evidence for claims given you have failed to do so.
Alpha3031··on Sticky wage norms and the real wage cost of unexpected inflation
Not many people buy GA aeroplanes, so it's a sensible decision for those to receive a lower weighting than other consumer goods in a general purpose index like the CPI.

There is a divergence between the prices of goods the US produces and those that consumers purchase, as measured by the GDP IPD and CPI respectively, but that is not something that is appropriate to measure using one highly idiosyncratic good, nor even a basket of goods like "things I personally want to purchase", if one is wishing to produce numbers that others agree are reflective of the average experience of many.

Alpha3031··on Sticky wage norms and the real wage cost of unexpected inflation
> The minimum wage is a strawman by comparison: it doesn't actually help workers.

[citation needed]

By measures such as own-wage elasticity [DUBE & ZIPPERER] increases to the minimum wage evidently increase wages more than they decrease employment.

[DUBE & ZIPPERER]: https://www.nber.org/papers/w32925

Alpha3031··on Sticky wage norms and the real wage cost of unexpected inflation
As I understand it, that's a fairly common definition for relative poverty (either that or 60%) though of course you're still right that it's best to confirm both the threshold the methodology is the same (or use an explicitly international comparison that covers both/all compared countries in the same work, such as those produced by the OECD or similar bodies).
Alpha3031··on Norway Should Buy OpenAI
Getting to parity with less money means that you can go further with the same amount of money.
Alpha3031··on AI;DR (AI; Didn't Read)
Again, you are free to interpret my comments however you wish. My primary intent was to express my belief that people having different preferences to you regarding reading AI output does not constitute some bizzare hivemind, which you appear to have accepted. Your preferences regarding communications between yourself and your wife are none of my business. I recognise preferences may exist, and I am not one to kink-shame.
Alpha3031··on AI;DR (AI; Didn't Read)
I said I would find the action bizarre ("that", not "them"). Personally, I don't find it unfortunate to admit that some things still surprise me. If you choose to interpret that as judgement of your wife as a person, then that is up to you.
Alpha3031··on AI;DR (AI; Didn't Read)
I've never seen anyone post a copy-and-paste of some pages of the Google search results while also excluding the search term, and I would find that bizzare. I would assume that you, like most people posting a Google search, would post the link to that search, which in the case of Google includes a copy of what would be equivalent to the prompt. In my experience, most people posting Google links would also include a short explanation of why they think it relevant were it not immediately obvious.
Alpha3031··on AI;DR (AI; Didn't Read)
I'm in the "post both and let your readers figure out which, if any, they prefer" camp but even if you didn't (and unless you were using one of those super special models that can't be released according to the companies) I would hope that any literate adult that thought it valuable would be capable of consulting something with comparable knowledge and experience.

Of course, given that posting one is not mutually exclusive with posting the other, I would still think that posting both is the easiest way to satisfy both people who prefer to read the prompt and those who prefer the read the output.

Alpha3031··on AI;DR (AI; Didn't Read)
The author giving the slightest amount of shit seems like a big assumption to make, but ultimately even it that were the case I don't see why I would prefer what I read to be curated in that specific manner (having access only to what has been filtered past some black box). As opposed to being able to choose what to read from the whole chat history.
Alpha3031··on AI;DR (AI; Didn't Read)
People having preferences different from you is not a hive mind.
Alpha3031··on Anthropic's 'watermark' text adulteration in Claude is a perversion of writing
Less unlikely, but I would still suggest that it is somewhat unlikely for any recent LLM to place any significant probably on, e.g. "disappointment" instead of "failure" as the appropriate token (or series thereof) after "kidney" or any similar example.
Alpha3031··on Anthropic's ‘watermark’ text adulteration in Claude is a perversion of writing
If it's not a problem, then there should be no issue with not concealing the fact, no? Lying about things one considers inconsequential is a useful signal about one's willingness to lie with little benefit.
Alpha3031··on Anthropic's ‘watermark’ text adulteration in Claude is a perversion of writing
Seems unlikely unless Anthropic invented a time machine, given that the phenomenon predated Claude 1 by two years, and their stated introduction of watermarking (August 2) by 5.
Alpha3031··on Guiding Ships with Moire Patterns
[delayed]
Alpha3031··on Grok 4.6
Supposedly people are getting ~40 tps decode at Q8 on 2× DGX Spark (higher for Q4) which is what I assume they're suggesting is just under $10K USD. Prefill is just above 1.5k so TTFT is maybe 2 to 4 minutes? I don't have two DGX Sparks myself so I can't confirm and not 100% sure if that number is with or without speculative decoding already in use (if not probably around 60 to 80 if enabled?).
← PreviousPage 2 of 9Next →