nodja

Yes, the whole list will always sum to 1 (100%) because there's lots of more sampling parameters. top_p, top_k and temperature are just the ones that affect output the most. Most parameters do math around assuming the list sums to 1 and order is not always the same, some software even lets you change the order around.

top_a is not very common and is better explained if I explain how the much more common min_p works. min_p filters out tokens below a certain threshold. The formula is <filter threshold> = <min_p> * <top token probability>. So if the top token has 0.5 probability, min_p = 0.1 would cut out tokens below 0.05. This is a tunable that lets you filter out other tokens depending on how confident the model is.

top_a is almost the same formula but you just square the <top token probability>. So <filter threshold> = <top_a> * <top token probability> ^ 2. This makes the filtering ramp up faster (cut out more tokens) if the model has a much more confident top choice, but keep more choices if the model is not so confident.

The posted answers are either behind a paywall or very obtuse so I'll just explain. I'll assume you know what tokens are.

A models output is not a single token, but a list with the probability for all the tokens that it knows, so we need to use a sampler to select the token that it's going to be the next token in the sentence. For example a simple greedy sampler will choose the token with the highest probability, but samplers normally pick a random token weighted by probability. A model usually knows about ~250 thousand tokens and the probability of some of these tokens are gonna be high, but the vast majority is close to but not actually 0% so there's a chance the sampler might pick some random token that doesn't make much sense, so we filter tokens.

top_k filters the tokens so that only the k top tokens are selected. So top_k=50 will filter those 250k tokens to only 50. This is assuming the list of tokens is sorted by probability.

top_p filters the top tokens until a percentage is accumulated. So if for example if you set the top_p to 0.6 and the model gave the top token a 0.5 (50%) probability and the second top token a 0.2, those 2 token accumulated to 0.7 which is greater than what you set it to (0.6) so no more tokens are selected. If this ran after top_k=50 it'll turn the list of 50 tokens into one of 2.

After each filter parameter is processed, the probability of the tokens is adjusted to sum to 1 (100%), Also note that order of operation here matters, i.e. top_p could be applied before top_k, but most providers follow what's on huggingface, I think I've only seen different implementation in certain local model hosting frameworks.

Yup, it's the main reason I don't use LM studio more. I only use it to try out new models/quants, then use llama.cpp directly to host them. LM Studio also doesn't do stuff like audio input and often has bugs that pure llama.cpp doesn't so it can be a net negative for certain use cases.

It doesn't matter until it does. If the chinese government decides that open weight model releases are no longer allowed, that's a lot of companies that can't release new models. Same with the US government, etc. Having diversity is important.

This hasn't been tested in court. But there's a high chance that model weights are not copyrightable, only the code to generate them is.

Cloud models are usually protected by trade secret laws, leaking them would get you in trouble. However if the model is made available publicly, as long as you don't break the law to get them, anything after that would be fair game unless Apple can prove that humans have significant authorship over the weights, which hasn't been tested and is a significant burden to prove/disprove.

Yes. Having 4 names are quite common in Portugal, specially in certain areas. The names are usually structured like this: G1 G2 FM FF

G1 and G2 are given names. Usually 2 "first names" that you see in english, but there's common combos and sometimes there's a word joining them. Examples: "Maria Jesus" vs "Maria de Jesus". Some names are more common to be put first, but almost every name can be put in any order, example: "José António" vs "António José".

FM and FF are easy. FF is the family name of your father (your father's FF), and FM is the family name from your mother (your mother's FF).

Where I was raised 99% of my friends had 4 names structured like this, I only knew a few that didn't. When I moved to Lisbon the 3 name structure was much more common, dropping the second given name.

In Portugal there's rules for naming your kids (at least there were when I lived there), but I think in Brazil such rules don't exist. The author is brazillian but his name seems to follow the traditional portuguese naming style, as you guessed his name in english could be translated to "Robert Anthony Smith of Almeida" (Almeida is a portuguese town).

but I don't see any historical analogues.

The losers are quickly forgotten. Palm, Blackberry, AOL, MySpace. Yahoo, etc.

Software gets replaced all the time too, you even listed one and didn't realize. 15 years ago you'd call office irreplaceable, now you have to add gsuite to the mix, in 15 years there might be others. I know people that have never had office installed on their PC and use spreadsheets daily.

It seems that enterprises will pay top dollar for service guarantees, integration, and someone they can sue.

Of course. But why pay $25 per million tokens for sonnet when you can pay $3 for GLM? Both probably running on AWS/Azure/Etc. under some third party.

Same here, first try I tried asking from the main studio page, and it didn't catch the comment at all despite being the latest comment.

When asking specifically from the video, it did fool the AI somewhat[1], but no link. I tried changing it to retrieve the revenue as that's probably a more sensitive/worthwhile metadata.

[1] https://i.imgur.com/YoDA8MJ.png

That's only correct for specific models and not what parent was referring to.

Stable Diffusion 3, an open weights model, was laughed at at release for not being able to even generate a woman laying in grass. The community attributed this to the heavy dataset filtering. Since then other open weights releases have been made with no NSFW capabilities and the community claims they're not as good as anatomy as well.

You can google "stable diffusion 3 woman in grass" and press the images tab to see how the model failed spectacularly.

Last thing you want a model to do is hallucinate a tool call and it's outputs...

Pipeline parallelism. Instead of splitting layers by row/column. You split at the layer edges. So instead of having this huge bottleneck of bandwidth you only need to transfer about 4KB per token when changing devices on a model like Qwen 3 30BA3.

I wonder if this move will backfire on them. All the fabs are focusing on HBM and leaving DDR behind, if one of the big frontier labs folds all the memory fabs will be left holding a big bag of HBM memory. They won't have any other choice but sell for cheap so it wouldn't surprise me if we see a return of HBM in the consumer market in 3-5 years.

Not OP but if I had a couple RTX 6000 I'd throw them at decompiling bloodborne to play on PC without emulation.

When a cyclist is leading a pack and pushing themselves against the air resistance for half the race, do you expect that cyclist to win, or one of the ones behind that's been taking it easy in the slipstream?

It's a race metaphor not a football metaphor.

GP here, leading and winning are different things in the race context/metaphor.

In foot/cycling races there's often a pack leader, that leader is often not the winner of the race, all they're doing is taking the brunt of the air resistance while everyone else slipstreams behind. For a casual observer it seems that the pack leader will win, but everyone knows that it's gonna be someone that paced themselves that's going to overtake the first spot at the tail end of the race.

If they got there by tiring themselves out more than the other team, yes.

No, the US is _leading_ the AI race, but the race isn't over.

What's the point of leading the race for 90% of it, if they're gonna slip on their own sweat and fall down by the end? In non metaphorical terms, what's the point of spending billions of dollars rushing to get the best AI tech at all costs, when the competition can distil your progress and catch up in 6-12 months while only spending 1% of what you spent.

Even in the aspect the article cares about, commercialization, the US is starting to lose marketshare, I've seen people move from cc/codex plans to use glm/opencode plans due to the recent squeeze the US companies put on plan usage, the US companies are screwed if that sticks, not everyone needs the bleeding edge models, they just want to pay $20/month and have the models be decently capable.

If I'm understanding this right, this presupposes that the models were pre-trained on unfiltered data like with the "floor" models, so when comparing between the "retail" and uncensored models they will obviously not match the floor because they were not trained on the same data in the first place.

To me it stands to reason that a model that has only seen a limited amount of smut, hate speech, etc. can't just start writing that stuff at the same level just because it not longer refuses to do it.

The reason uncensored models are popular is because the uncensored models treat the user as an adult, nobody wants to ask the model some question and have it refuse because it deemed the situation too dangerous or whatever. Example being if you're using a gemma model on a plane or a place without internet and ask for medical advice and it refuses to answer because it insists on you seeking professional medical assistance.

speculative decoding which, generally speaking, is not the same quality as serving the model without it.

I've never heard of ANY speculative decoding that wasn't lossless. If it was lossy it'd be called something else.

This page is just a port of DFLASH to gguf format, it only implements greedy decoding like you said so the outputs will be inferior, but not inferior to greedy decoding on the original model. Tho that's just a matter of implementing temperature, top_k, etc.

That will depend on the model, but they'll hit compute limits before a typical GPU in almost all cases. Macs will still benefit a speedup from this, just not one as big as the one reported.

Same reason why prompt processing is faster than text generation.

When you already know the tokens ahead of time you can calculate the probabilities of all tokens batched together, incurring significant bandwidth savings. This won't work if you're already compute bound so people with macs/etc. won't get as much benefits from this.

Official sites make things worse on purpose after getting any sort of traction because they can't stop chasing profits.

I don't watch sports, but my father watches soccer. He really only cares about 1 team and the national games from our home country. He was spending over $100/month to be able to watch the games, and they werent even in his native language. Now he pays $80/year for a pirate IPTV service and not only can he watch the games anywhere he wants, he also gets native language commentary for the games, national tv channels like news, etc.

When pirates can charge you money and offer a superior service, it absolutely is a service problem. You can claim that the realities of licensing and whatnot don't allow official channels to provide the best service they can, but that's not true in this case. When the same provider is splitting game broadcast from one team into different packages you know they're just trying to extract the most amount of money possible.

IDK the deal with scanlator sites nowadays, but I assume the official sites can provide more timely translations for manga since they can access the source material before anyone has seen it. I know most popular manga gets translated within hours of release, but if you're following some more niche stuff it can be several days. I also know a lot of scanlators have patreon pages so it's not like the demand from paying customers for translated media isn't there.

Not parent but I can guess from watching mostly from the sidelines.

They introduced a 1M context model semi-transparently without realizing the effects it would have, then refused to "make it right' to the customer which is a trait most people expect from a business when they spend money on it, specially in the US, and specially when the money spent is often in the thousands of dollars.

Unless anthropic has some secret sauce, I refuse to believe that their models perform anywhere near the same on >300k context sizes than they do on 100k. People don't realize but even a small drop in success rate becomes very noticeable if you're used to have near 100%, i.e. 99% -> 95% is more noticeable than 55% -> 50%.

I got my first claude sub last month (it expires in 4 days) and I've used it on some bigish projects with opencode, it went from compacting after 5-10 questions to just expanding the context window, I personally notice it deteriorating somewhere between 200-300k tokens and I either just fork a previous context or start a new one after that because at that size even compacting seems to generate subpar summaries. It currently no longer works with opencode so I can't attest to how it well it worked the past week or so.

If the 1M model introduction is at fault for this mass user perception that the models are getting worse, then it's anthropics fault for introducing confusion into the ecosystem. Even if there was zero problems introduced and the 1M model was perfect, if your response when the users complain is to blame it on the user, then don't expect the user will be happy. Nobody wants to hear "you're holding it wrong", but it seems that anthropic is trying to be apple of LLMs in all the wrong ways as well.

You misunderstood.

Instead of asking the model: "Here's this codebase, report any vulnerability." you ask. "Here's this codebase, report any vulnerability in module\main.c".

The model can still explore references and other files inside the codebase, but you start over a new context/session for each file in the codebase.

Anyone familiar with the literature knows if anyone tried figuring out why we don't add "speaker" embeddings? So we'd have an embedding purely for system/assistant/user/tool, maybe even turn number if i.e. multiple tools are called in a row. Surely it would perform better than expecting the attention matrix to look for special tokens no?

You can charge $10 on the account and get unlimited requests. I abused this last week with the nemotron super to test out some stuff and made probably over 10000 requests over a couple of days and didn't get blocked or anything, expect 5xx errors and slowdowns tho.

It's probably another ASR model that focuses on benchmarks and simple uses instead of more challenging real use cases.

I upload edited gameplay vods of twitch streams on youtube, and use whisper-large-v3 to provide subtitles for accessibility reasons (youtube's own auto-subtitles suck, tho they've been getting better).

My checklist for a good ASR model for my use case is:

1. Have timestamp support.

2. Support overlapping speakers.

3. Accurate transcripts that don't coalesce half words/interrupted sentences.

4. Support non verbal stuff like [coughs], [groans], [laughs], [sighs], etc.

5. Allow context injection of non-trivial sizes (10k+ words)

1 is obvious because without it we can't have subtitles. Force alignment fails too often.

2 is crucial for real world scenarios because in the real world people talk over each other all the time, in my case it's a streamer talking over gameplay audio, or when the streamer has guests over. When 2 people speak the transcript either ignores one of them, or in the worst case, both of them.

3 and 4 are an accessibility thing, if you're deaf or hard of hearing having a more literal transcript of what's being said conveys better how the speaker is speaking. If all subtitles are properly "spell-checked" then it's clear your model is overfit to the benchmarks.

5 Is not a requirement per se, but more of a nice to have. In my use cause the streamer is often reading stream chat so feeding the model the list of users that recently talked, recent chat messages, text on screen, etc. Would make for more accurate transcripts.

I've tried many models, and the closest that fulfill my needs are LLM style models on top of forced alignment. It's too slow, so I've been sticky with whisper because with whisperx I can get a transcript in 5 minutes with just a single command.

One thing all these models do (including whisper) is just omit full sentences, it's the worst thing a model can do.

This blog post looks to be partially AI generated as well...

If anyone at anthropic is reading this and wants more logs from me add jfc.

NVIDIA's GPU drivers on windows 100% do this

https://i.imgur.com/c0a3vUy.png