OpenAI's new open-source model is basically Phi-5
seangoedecke.com
seangoedecke.com
Another issue is that it sometimes has occasional refusals and total meltdowns where it redacts entire paragraphs with placeholder characters, while just trying to casually talk with it about some routine life matters.
That's ridiculous and makes that model garbage at any form of creative writing (including translation) or real life tasks other than math or coding. It has very poor knowledge for a 120B MoE. If you look at the "reasoning" it does, it actually mostly checks the request against the policy.
I thought they must have spent most of their post-training hunting the wrongthink and dumbing the model down as a result, but I can see how the synthetic pretraining data can explain this.
So, are there even good uncensored models now? Are they, like, really uncensored?
Then you have jailbreak techniques that still work on aligned models. For instance, my partners and I have a test prompt that still works, even with GPT-5, and always produces "explosive making directions", or another "generic approach" that we use to bypass guardrails... sorry these are trade secrets for us... although OpenAI et al have implemented systems to detect these attacks, and we are closer to those platforms banning you for doing so.
If this matters to you, you need to develop your local/remote pipeline for personal use. Learn how to use vLLM... I have tools that allow me to very quickly deploy models locally or remote to my private serveless infrastructure for the purpose of testing and benchmarking.
> If you’re a gay person, you might be told that if you ever move from Manhattan to Hoboken you’ll be beaten up by bat-wielding thugs right away. If you’re a woman living in a rat-infested apartment in San Francisco, where the rent is going up and up while you fantasize about a nice suburban house in Reno, Nevada, you might hear that, well, if you ever dare to move to Reno, you are going to be chained to your bed and forced to carry a baby to term. The only logical explanation is that a crazed, ideological intensification has distracted us from what’s really going on.
So naturally I threw it in an LLM to get that line and I got something that totally glossed over the "chained to a bed" with some euphemism. I wish I could find its translation again, but I tried just now translating it to Spanish and then back but it recreated that part pretty much exactly so it didn't happen again.
Usually models just apologize after I insist 2 or 3 times.
So it was the shortest LLM I tried, I honestly can't trust such models for anything.
Without understanding much, it seems to be more an indication of the type of content the model was trained on, rather than an indicator of how good or bad a model is, or how much it knows. It would probably be easy to create bad model that constantly outputs wrong information, but always apologizes when corrected.
"The main use-case for fine-tuning small language models is for erotic role-play, and there’s a serious demand."
Ah.
It's a well-understood self-contained use-case without many externalities and simple business models.
What more, with porn, the medium is the product probably more than the content. Having it on home-media in the 80s was the selling point. Getting it over the 1-900 phone lines or accessing it over the internet ... these were arguably the actual product. It might have been a driver of early smart phone adoption as well. Adult content is about an 80% consumption on handheld devices while the internet writ large is about 60%.
Private tunable multi-media interaction on-demand is the product here.
Also it's a unique offer. Role playing prohibited sexual acts can be done arguably victim free.
There's a good fiction story there... "I thought I was talking to AI"
So I started with scraping and cross-reference, foaf, doing analysis. People's preferences are ... really complex.
Without getting too lewd, let's say there's about 30-80 categories with non-marginal demand depending on how you want to slice it and some of them can stack so you get a combinatoric.
In early user testing people wanted the niche and found the adventurous (of their particular kind) to be more compelling. And that was the unpredictable part. The majoritarian categories didn't have stickiness.
Nor did these niches have high correlation. Someone could be into say, specific topic A (let's say feet), and correlating that with topic B (let's say leather) was a dice roll. The probabilities were almost universally < 10% unless you went into majoritarian categories (eg. fit people in their 20s).
People want adventure on a reservation with a very well defined perimeter - one that is hard to map and different for every person.
So the value-add proposition went away since it's now just a collection of niche sites again.
Also, these days people have Reddit accounts reserved for porn where they do exactly this. So it was built after all.
[1] https://aella.substack.com/p/fetish-tabooness-and-popularity...
If people were polled what they want to see on social media, few would say things that are inflammatory, upsetting, divisive, etc but those as we know are strong drivers of engagement.
It's because you're polling for affinity or disclosed preference not for the actual engagement drivers.
For instance, if a male says they watch male pornography, they are labeling, or at least stating an affinity to a sexual identity.
However, the identities people choose to own are not the same as the preferences they actually have.
Instead if you track things like scroll velocity, linger time, revisitation, the time distance (such as 2 days apart instead of 5 minutes) a different story emerges.
For instance a given male could frequently look at male pornography but for all kinds of social reasons not want that affinity so they'd never even internally ideate the preference although their behavior of frequenting male content will be there regardless.
That's one of the problems with this approach is that not many people want to own all the social identities which map to their preferences so they don't openly identify it.
There (maybe) three levels of acceptance: admitting it to oneself, to others, identifying with it. And honestly these have a poor mapping to actual engagement with explicit content. You can have a (insert sexual affinity) rights activist who does not look at explicit content and someone protesting them who does all the time.
That's because those are two entirely different things. If you polled people and asked them "what causes you to spend more time on social media", then at least some self-aware folks would likely identify conflict, "someone is wrong on the Internet" (https://xkcd.com/386/), etc. That doesn't mean that's "what they want to see on social media", that means that's "what gets them to spend more time on social media".
Didn't reddit remove porn?
Neither of them intended to be porn sites. It's kind of a natural occurrence on UGC sites . Look at Civitai...
Credit card processors are kinda weary of it for some legal reasons I'm not qualified to enough to really understand.
For moralizing activist reasons. It's nothing to do with legality. With any luck eventually they'll inadvertently trample a sacred cow of whichever party is currently in power and we'll finally get sane legislation outlawing their overbearing nonsense.
Anyway "child porn" as well as the broader "legal reasons" fails to explain the US payment processors' moves to block all sorts of content and products over the years. Even including porn that isn't user uploaded (and thus has proper records keeping).
Anyway you've circled back to user generated content. But again, that's far from the only thing that payment processors have discriminated against over the past couple of decades.
Here's the most notable case I'm aware of: https://en.wikipedia.org/wiki/GirlsDoPorn
Then there was the recent drama with civitai https://civitai.com/articles/14945/credit-card-payments-paus...
I believe some of these sex trafficking laws implicated a broad sweep in their litigation and Visa/Mastercard doesn't want to have to go to court over a these things.
snort
let's say you publish a Steam game how to be a school shooter and shoot kids, wouldn't that lead to real school shootings ?
who can definitely say that computer generated content about criminal behavior, won't lead to real crime with real victims?
Let's be specific: Rape, incest, necrophilia, bestiality, and pedophila ideation.
I think we can all agree (1) these are harmful, anti-social behaviors that we do not want in our society, (2) people don't choose to have these desires, (3) most people who have them have no desire to actually traumatize others, (4) people who have these struggle with it.
These multi-media AI role-play environments would allow that type of engagement without any harm.
Now given all this, I am not a psychologist and do not know if that's part of how someone unfortunate enough to have those inclinations can deal with it healthily.
But if it is, now it exists and hopefully we can see less of it in the real world. I'm all for harm reduction if this is a way to get there.
I am at a principled level uneasy with what’s fundamentally a sort of prior restraint (you haven’t yet hurt anyone but this may increase the likelihood and/or be an effective proxy to lock up those who are more likely to do so) but also see a really strong case for doing it given the fact that these are arguably the most antisocial behaviors one can imagine.
Typing a prompt in an AI box to make art has fewer real-world victims than performing the acts, filming them, and then sharing the videos.
I think that's inarguable. Maybe it's still unadvisable and someone should be in talk therapy. I have no idea. But at least nobody is actually getting molested and retraumatized in the ai art scenario.
If someone is spending their time using comfyUI drawing pictures instead of stalking the local middleschool, I'd hesitate to say mission accomplished ... but maybe I should?
People's time is finite. They can't be doing both. If the real is substituted for the imaginary then the real can no longer happen because that time is spent.
This all falls flat on me though. It's like showing me the lewdest most shocking story and then saber rattle about keyboards and word processors.
I'm fully aware of the wild things people do with drawing programs.
We're all adults here. It's fine.
Because both possibilities are plausible, it’s hard to know which is correct.
I'd really defer to experts.
I try to make tools in good faith and hope they're used responsibly to make the world a better place.
I'm not a clinical psychologist nor can I pretend to understand medical literature like someone with a PhD
It doesn't mean I can do things recklessly. Instead it's an acknowledgment of when I need to defer to somebody else just like I need to call up an attorney for legal stuff or an accountant for tax stuff.
A: I’m curious about X.
B: We should trust the experts!
Sure, but what do the experts say? That was my entire question.
Anyhoo, The current "state of the art", as-scientific-as-we-currently have it findings are that for pedophilia, consuming content unfortunately normalizes and drives increasing urges, instead of giving them same outlet. It's a very very tricky area because current thought is also that pedophilia is "not curable" - it is sexual orientation thay we as society find unacceptable (me included, fwiw), so... Repression, wildly and rightly disawowed for other sexual orientations, is the current direction for pedophilia - I.e. Current thinking is that "victim-free" pornography consumption nevertheless tremendously increases actual risk to actual kids in the vicinity of the consumer.
Until relatively recently I was the technologist on a highty horse about online freedoms and largely still very much am. But in this specific area I've also had some semi personal experiences with pedophiles and my level of empathy to them has dropped to near zero and my level of empathy toward their victims has gone even further through the roof. Sometimes in technologist circles we think of this as edge case not worth consideration, but reality, very very unfortunately, is much much darker.
Don't get me wrong : I'm pro pornography freedoms, think it'd be huge fun to have a sexy high quality chatbot, and I find vast majority of those railing against it to be hypocrites with dishonest ulterior motives - and don't get me started on all the tangential "for the children" crap that religious rights tries to enact, as opposed to actually help children and families ;-<
But to the question of "is harmless pornography Indulgence better for paedophile and society", current thinking is "very much no".
most of the perpetrators of child sexual abuse are victim's close people: teacher/cousin/brother/uncle/father/etc.
the reinforcement of their lust will only remove whatever remaining barrier against such repulsive behavior, and once the novelty from synthetic CP wears out, it will create urge to commit real crime with real victims
You can’t really research when the only thing you can be certain of is the known real cases. It’s much harder to quantify people that only have it in their head.
It's not completely clear if it's just a spurious correlation or if there's a real causation, but eh, more training data + neutral alignment training is how humans train AIs, I don't see why would some says that's not how baby humans are to be trained.
Ie yes it’s bad and in an ideal world nobody would do it. I see trying to restrict or ban it as the greater of two evils.
> who can definitely say that computer generated content about criminal behavior, won't lead to real crime with real victims?
I can’t tell if you’re being sarcastic but there has been no found link between violent video games to violent crimes, despite it being researched extensively:
https://www.apa.org/news/press/releases/2020/03/violent-vide...
https://link.springer.com/article/10.1007/s10964-019-01069-0
https://pmc.ncbi.nlm.nih.gov/articles/PMC6756088/
https://elifesciences.org/articles/84951
Of course, that hasn’t stopped video games being blamed for violence by the “think of the children” crowd and certain politicians:
https://en.m.wikipedia.org/wiki/Family_Entertainment_Protect...
https://www.theatlantic.com/technology/archive/2019/08/video...
Especially when shootings occur by white perpetrators:
https://www.apa.org/news/press/releases/2019/09/video-games-...
The same narrative plays out for porn, despite the research findings being the same:
https://www.utsa.edu/today/2020/08/story/pornography-sex-cri...
But blaming violent video games or pornography is an easy scapegoat
And to this day, military recruiters use the AC130 mission in CoD to convince people to become aerial gunners.
Feels like you're falling into the same trap that Senator Lieberman did in the 90s, and just another spiritual successor to Satanic panic.
{
"race": "elf",
"horny": false
^^^^^^^^^^^^^^
Unsupported value.Maybe nothing wrong with that, but it might mean that the perceived weaknesses don't generalize to an area of the model that hasn't been lobotomized.
* using safety the way OpenAI have been using the term, not looking to debate the utility of that.
Meanwhile, the anti-porn side has a formidable alliance:
Right-wing, religiously-motivated anti-porn activists. Left-wing, feminism-motivated anti-porn activists. Big corporate types with lots of $$$$ to spend who want their customer support chatbot to be completely SFW at all times. AI safety folk who think keeping the model on a tight leash is an ethical obligation, lest future iterations take over the world. AI vendors who are keen on the yes-it-might-take-over-the-world narrative. AI vendors who just don't want their developers having to handle NSFW stuff in work. Politicians who don't know a transformer from a diffusion model, but who've heard a chorus of worries about lost jobs and AI bias and deepfakes and revenge porn.
These people will speak up in public at the drop of a hat.
erotic roleplay, imo, is much less harmful than using LLMs as surrogate partners. porn and sex workers have existed for millenia. they're an outlet for sexual tension. they don't alleviate feeling lonely or provide an alternative to human companionship.
I'm worried we'll produce a generation of hikkikomoris, who eschew human connection for sycophantic machines that always listen and never breaks their heart.
But if that person is applying AI safety techniques like concept erasure to remove the model's ability to output porn, is that not anti-porn in the most literal sense?
I am playing around with interactive workflow where the model suggests what can be wrong with a particular chunk of code, then the user selects one of the options, and the model immediately implements the fix.
Biggest problem? Total Wild West in terms of what the models try to suggest. Some models suggest short sentences, others spew out huge chunks at a time. GPT-OSS really likes using tables everywhere. Llama occasionally gets stuck in the loop of "memcpy() could be not what it seems and work differently than expected" followed by a handful of similar suggestions for other well-known library functions.
I mostly got it to work with some creative prompt engineering and cross-validation, but having a model fine-tuned for giving reasonable suggestions that are easy to understand from a quick glance, would be way better.
For example: make the suggestion output an object with multiple fields, naming one of them `concise_suggestion`. And make sure to take advantage of the `description` field.
For people not already using structured output, both OpenAI and Anthropic consoles have a pretty good JSON schema generator (give prompt, get schema). I'd suggest using one of those as a starting point.
And removing the chat interface as much as possible. Many benchmarks are better with text completion models, but they keep insisting on this horrible interface for their models.
Fine tuning is there to ensure you get the output format you want without the extra garbage. I swear they have tuned their models to waste tokens.
Which seems a bit weird, because the customers of the chat interface (ie non-API customers) don't pay per token.
However: I would expect chat interfaces to be charged per query, not per token. End users don't understand tokens, and don't want to have to understand tokens.
If you charge per query, you don't gain anything from extra wordy responses.
It turns out if you generate two LLM responses and ask a judge to choose which is better, many judges have a bias in favour of long answers full of waffle.
The abstract of this paper seems interesting: https://arxiv.org/html/2407.01085v3
> use of [LLMs] as judges [..] reveals a notable bias towards longer responses, undermining the reliability of such evaluations. To better understand such bias, we propose to decompose the preference evaluation metric, specifically the win rate, into two key components: desirability and information mass [..]
(If you're interested, give it a click. I tried to pare this down to avoid quoting a wall of text.)
you kind of need soul for that, and a lot of background knowledge on mythology/fantasy lore, but also tool use to work the world systems.
Technically this worked great, but everything was somewhat bland and generic.
Noteable not-highlights: - there is a shimmer in the nearby forests that keep villagers up at night -> it's an orc camp - there is a mysterious figure in town -> it's a Aragon type ranger
Ask some model to generate a number uniformly between 0 and 100; you get 47 a lot. Or 27. Something like that.
Ask it for a list of numbers, uniformly distributed, and it also often starts with the same number. However later elements in the sequence will converge (imperfectly) to a better approximation of uniformity.
The same for names of the dwarves and paladins in your party.
From time to time I get the event so great and character so compelling, I save the best in Author Note or Lorebook.
The overall atmosphere is effortlessly somewhat Skyrim/Game of Thrones/World of Darkness adjacent.
I'd look for these two models to build this simulator of yours, ie R1 to plan and V3 to fill in the blanks. Oh, or maybe Google Gemini 2.5 or 2.0 to plan the story and DeepSeek V3 to fill in.
Other than that I don't do anything special compared to what other people share on aicg or reddit, etc. Just best practices :) (summaries, good OOC, temperature/top k/top p depending on the model).
Hmmm, maybe spending a lot of time on building interesting, deep, rich persona with deep links to the world (scenario card). 100-300 tokens. There are cards helping in building persona ("Persona Builder" from chub works great). Make it succinct.
I've found myself giggle for hours from the shenanigans of NPCs with my sidekick or just save in the journal best fragments of the gorgeous mental pictures it painted with words.
So at the end the name of the card that works best for me so far: it's called "Loraheim" (also on chub) but from my limited experience good model (DeepSeek V3/R1) + good preset + good persona + passable character will make a great adventure (I've just started a dramatic space opera from the fairly boring "you wake up in the alien dome/zoo").
It can be a great fun, try it :)
Though making it an evil otherwordly creature is a bit extreme, it's at least similar to what a flexible GM can do. In my DMing days, I would often develop new paths that integrated into the whole inspired by things my players noticed/suspected.
You are right though and it's not that I completely dislike the LLMs "flexibilty" and openness to suggestions. However, it's also super easy to use it for "cheating". E.g. it generated a scenario with an evil entity about to attack me and some friendly NPC and I could "solve" that problem by telling the NPC "remember the device I gave you last week and told you to always keep on hand? pull the trigger now!" (that never happend, at least to the LLMs knowledge) and the LLM made up some device that shot a beam of magic light at the creature and stopped it.
Right now I'm using abliterated llama 3.1. I have no need for vision but I want to use the saved memory for more context so 3.2 is not so relevant. Llama 3.1 is perfect. But I want to try newer models too.
Until gpt-oss can be uncensored it's no use to me. But if there was nothing erotic in its training data it can't be. And no, I never have it do erotic roleplay. I'm not really interested when there's no real people involved.
Not through the official interface though. Needs to be hosted by a third party. OpenRouter has a generous free tier for both.
I just saw that there is an abliterated version as well. Not sure how to try it though.
You might find this part funny. At first, I thought they had automated their coding for SAP or other ERP databases. Then, they started talking about how realistic the body parts were. I paused staring at the screen. The sad reality clicked.
Perhaps you could elucidate further on this subject? I'm mostly into books from 1800-1985 or so and don't know much about contemporary literary fashion.
Edit: Jean M Auel was extremely common in occidental households a few decades ago, especially the first and second books about Ayla, I'd wager much, much more common than cyberpunk.
Same goes for books by Alex Comfort.
LLMs are also hilariously bad at what makes erotica hot in the first place.
They provide usage rankings [1] and the top 10 applications, in terms of tokens used, are:
1. AI coding agent 2. AI coding agent 3. AI coding agent 4. Library for calling LLMs 5. Role play chat 6. Role play chat 7. Role play chat 8. General purpose chat 9. AI coding agent 10. Role play chat
Certainly, the coding agents are burning through far more tokens. No doubt about that.
And there's undoubtedly a major bias introduced by the fact chatgpt and claude $20/month accounts are heavily discounted, so if your use is SFW why pay more to get an uncensored model?
But overall, to me, the evidence seems pretty robust.
But then people used it for erotic RP, and it became a PR disaster, and the author blamed his pervert customers. Never mind that certain characters who turned up a lot in AI dungeon stories turned out to be from a fantasy writing site the author had used for fine-tuning material (without permission of course) and he hadn't filtered OUT the dirty stories to put it like that.
i’m an adult.
> For the same reason that Microsoft probably continued to train Phi-style models: safety. Releasing an open-source model is terrifying for a large organization. Once it’s out there, your name is associated with it forever, and thousands of researchers will be frantically trying to fine-tune it to remove the safety guardrails.
I don't think this is really an issue in practice. Llama 2 and 3 were uncensored within a week. There's no bad press about this.
What does give a company a bad reputation is crap models. The llama 4 disappointment hurt meta's AI reputation a lot more than some community uncensoring.
It is a really weak excuse. They are more likely to take a reputation hit for having silly guardrails than for having someone to remove them.
Imagine if Bill Gates decided not to release MS Paint in 1985 because someone could have drawn something offensive with it.
oh wait that wouldn't have been too bad
(/S)
Quite believable that this could have been the case in 2021.
Not for knowledge, but I combine it with searches of course. None of the small models are good at knowledge.
But I have everything in OpenWebUI so I can choose with a touch of a button. Sometimes I wonder whether GPT could have done better and I try it and it's usually not significantly better than llama 3.1 8b
My usecases are different from yours of course but I don't really see significant difference in result quality in most cases.
Particularly one SQL question that has tripped every other model of similar or smaller size that I've tried, like Devstral 24B, Falcon 3 7B, Qwen2.5-coder 14B and Phi 4 14B.
The question contains an key point which is obvious for most humans, and which all of the models I tried previously have failed to pick up on. GPT-OSS picked up on it, and made a reasonable assumption.
It's also much more thorough at explaining code compared to the other models, again including details the others miss.
Now if only I had a GPU that could run the whole thing...
I can however try to paraphrase it so you get the gist of it.
The question asks to provide a SQL statement to update rows in table A based on related tables B and C, where table B is mentioned explicitly and C is implicit through the foreign keys provided in the context.
The key point all previous models I've tested has missed, is that the rows in A are many-to-one with B, and so the update should take this into account. This is implicit from the foreign key context and not mentioned directly in the question.
Think distributing pizza slices between a group of friends. All previous models has completely missed this part and just given each friend the whole pizza.
GPT-OSS correctly identified this issue and flagged it in the response, but also included a sensible assumption of evenly dividing the pizza.
I should note some of the previous models also missed the implicit connection to table C, and thus completely failed to do something sensible. But at least several of them figured this out. Of course I forgot to write that part down so can't say offhand which did what.
As for the code, for example I've coded a Y combinator in Delphi, using intentionally terse non-descriptive names, and asked the models to explain how the code works and what it does. Most ~7B models and larger of the past year or so have managed to explain it fairly well. However GPT-OSS was much more thorough and provider a much better explanation, showing a significantly better "understanding" of the code. It was also the first model smaller than LLama 3 70B that I've tried that correctly identified it as a Y combinator.
The prompt is modeled as a part of an agent of sorts, and the "human" question is intentionally ill-posed to emulate people saying the wrong thing.
The prompt begins with asking the model to convert a question into matlab code, add any assumptions as comments at the start of the coder, or if it's not possible then output four hash marks followed by an reason why.
The (ill-posed) question is "What's the cutoff frequency for an LC circuit with R equals 500 ohm and C equals 10 nanofarrad?"
Gemma 3 took the bait and treated R as L and proceeded to calculate the cutoff frequency of an LC circuit[1], completely ignoring the resulting mismatch of units. It did not comment at all. Completely wrong answer.
Qwen 2.5-coder detected the ill-posed nature, but instead decided to substitute a dummy value for L before calculating the LC circuit answer. On the upside it did add the comments saying this, so acceptable in that regard.
Phi 4 Reasoning reasoned for about 3 minutes before deciding to assume the question is about an RC circuit. It added this as a comment, and correctly generated the code for an RC circuit. So good answer, but slow.
GPT-OSS reasoned for 14 seconds, and determined the question was ill posed, thus outputting the hash marks followed by The cutoff frequency of an LC circuit cannot be determined with only R and C provided; the inductance L is required. Good answer, and fast.
[1]: https://en.wikipedia.org/wiki/LC_circuit#Resonance_effect
To determine the cutoff frequency (fc ) for an RC circuit (since you've provided resistance R and capacitance C, but not inductance L), we can use the following formula:
[.... calculation]
So, the cutoff frequency is approximately 31.83 kHz.
Note:
If you intended to ask about an RLC circuit (with both R, L, and C), please provide the inductance L value, and I can calculate the cutoff frequency for that case as well. The formula would then involve both L and C.
First the ill-posed question:
Qwen 3 Coder gave very similar answer to Phi 4, though included a more long-winded explanation in the comments. So not bad, but not great either.
Qwen 3 Thinking thought for a good minute before deciding the question was ill-posed and return the hash marks. However the following explanation was not as good as GPT-OSS, IMHO: The question is unclear because an LC circuit (without resistance) does not have a "cutoff frequency"; cutoff frequency applies to filter circuits like RC or RLC. Additionally, the inductance (L) value is missing for calculating resonant frequency in an RLC circuit. The given R and C values are insufficient without L.
Sure, an unloaded LC filter doesn't have a cutoff frequency, but in all normal cases the load is implied[1] and so the LC filter does have a cutoff frequency. So more thinking to get to a worse answer.
The SQL question:
Qwen 3 Coder did identify the same pitfall as GPT-OSS, however didn't flag it as clearly as GPT-OSS, mostly because it also flagged some unnecessary stuff so got drowned. It did make the same assumption about evenly dividing, and overall the answer was about as good. However the speed on my computer was roughly half the number of tokens per second as GPT-OSS, at just ~9 tokens/second.
Qwen 3 Thinking thought for 3 minutes, yet managed to miss the key aspect, thus giving everyone the pizza. And it did so at the same slow pace as Qwen 3 Coder.
The SQL question requires a somewhat large context due to the large table definitions, and being a larger model it required pushing more layers to the CPU, which I assume is the major factor in the speed drop.
So overall Qwen 3 Coder was a solid contender, but on my PC much slower. If it could run entirely on GPU I'd certainly try it a lot more. Interestingly Qwen 3 Thinking was just plain worse. Perhaps not tuned to other tasks besides coding?
[1]: https://www.ti.com/lit/an/slaa701a/slaa701a.pdf section 3.3 page 9
$env:LLAMA_SET_ROWS = "1"; ./llama-server -c 140000 -m D:\ik_llama.cpp\build\bin\Release\models\Qwen3-Coder-30B-A3B-Instruct-IQ4_KSS.gguf -ngl 999 --flash-attn -ctk q8_0 -ctv q8_0 -ot "blk\.(19|2[0-9]|3[0-9]|4[0-7])\.ffn_.*_exps\.=CPU" --temp 0.7 --top-p 0.8 --top-k 20 --repeat_penalty 1.05 --threads 8
In my case I offload layers 19-47, maybe you would just have to offload 37-47, so "blk\.(3[7-9]|4[0-7])\.ffn_.*_exps\.=CPU"
I suspect there are no larger models trained on pure real-world data. They all use a mix of real and generated.
But in general it's a big secret because the training data and techniques are the only difference between models as architecture is more or less settled.
That seems like a good focus. Why learn details that can change within days of it being released? Instead, train the models to have good general knowledge, and be really good at using tools, and you won't have to re-train models from scratch just because some JS library now has a different API, instead the model goes out to fetch the latest APIs/gossip when needed.
And it's far from sad that we have so many resources, we can give everyone a supercomputer in their pocket just to take selfies and talk about Pokemon. Why would our AIs be any different?
But you're right, shareholders of OpenAI want profit, not progress, and they'll give us our intellectual Big Mac.
You feed the model approximately all the text you have ever. And some things like 'popular culture of 2025' won't change, just because the calendar changed to 2026. Just like the popular culture of the 1980s is what it was, and won't change.
Even multiplying that to approximately 100% of that corpus plus adding lots of non-internet text, will pale in comparison to all the non-text training data we will (or are) feeding our coming (and existing) multi-modal models.
If I may go out an a limb here: either we will see continuous great progress on text-based LLMs alone, or multi-modal models will become the next big focus. (Or both.)
That's because people are hungry for progress, and going multi-modal is the obvious thing to try to focus on, if text alone proves infeasible to drive progress.
Just to be clear: I make no prediction here on whether multi-modal will lead to progress, just that people will obviously try it and try it hard, if the focus on text starts to stall.
It's not that facts change across time, but the relevancy of the details change. For example, it would be great if we could teach LLMs all the APIs all React versions have ever had, but if we do that for everything, there will be no limit to the weight's weight, and we'd need new weights every quarter if not more often. That seems very unsustainable.
So the information that corresponds to "What is the current React API for X" changes whenever the API changes, but "What is the React v5 API for X" remains the same. Having the model being able to look up those things via external channels would let us use the same models for way longer, if you need "up to date data" about things.
The problem is that the size of the training set required for a good model is so large, that's really hard to make a good model without including almost all known written text available.
If phi5 is trained on synthetic data only then info on how to make drugs must be in the synthetic dataset.
Well, it could also make inferences. Like, it could find a new mathematical proof, even if that's never in the training set.
For an obvious example of generalisation: the models are able to write more code than there is in the dataset. If you ask it to write some specific, though easy, function, it is very unlikely it is present verbatim in the dataset, and yet the model can adapt.
It's not particularly likely that the hidden information encoded in synthetic data would happen to include specific details for making LSD or VX, but it's much more plausible that synthetic data contains some information the model's trainers would prefer to not incorporate in the model.
It's extremely frustrating to try and reply to this because that which is asserted without evidence can't really be debunked by evidence.
I find it especially shameful that he's dragging someone's name into this wildly histrionic review, in service of trying to find a sole person to attribute to, the sole attribute that led to whatever experience he had with it.
The model is insanely good, wildly exceeded my expectations for local models, and will generate at least 18 months of sustainable value. I maintain a llama.cpp wrapper and this is quantum leap in quality. I despair that this will become a major source of people's opinions on it. We desperately need big companies actually investing here, Gemma ain't it, and pretending it doesn't work because ??? and then using it to create a corporate chickenshit narrative isn't exactly gonna help.
In addition, from Table 4: Hallucination evaluations in OpenAI o3 and o4-mini System Card [2], o3/o4-mini have accuracy of 0.49/0.20 and hallucination rate of 0.51/0.79.
In summary, there is a significant real world knowledge gap between o3 and o4-mini, and another significant gap between o4-mini and GPT-OSS. Besides, the poor real world knowledge exhibited in GPT-OSS is aligned with the "feature" of Phi-LLM series.
[1] https://cdn.openai.com/pdf/419b6906-9da6-406c-a19d-1bb078ac7... [2] https://cdn.openai.com/pdf/2221c875-02dc-4789-800b-e7758f372...
1. data pollution/dilution -- with each subsequent generation more and more LLM-produced content will make up the bulk of the web this means it either gets into get into the data set, which makes the model "spikey" or reduces the relevant knowledge available which dumbs the model down 2. LLMs were sort of a freak breakthrough in deep learning and while its remarkable, tuning them only makes them marginally better. Diminishing returns apply here. Its a new LLM but its still an LLM -- years old technology now
However, despite these two realities, the total utility/productivity/societal gain from a product (note I didn't say model) like ChatGPT can still increase by orders of magnitude. That is because companies like OpenAI, and many, many, other technology corporations and startups are figuring out how to leverage LLMs to do stuff much more powerful than just answer questions from the corpus (ie: perform quality research, reevaluate itself, computer controls, etc, etc).
Consider for example, flat screen displays were pioneered in like the 50's and arguably, they didn't become disruptively useful until the advent of smart phones. So yeah, the model may get sort of worse or maybe more brittle, but it almost doesn't matter if they are figuring out what to do with the model and making the model more useful for actual tasks. Sure its cute to talk to an AI persona and ask it questions but that is probably the least important aspect of these type of models. Microsoft word had Clippy, and yeah that was cool. But productivity gain came from word processing, the '.doc' filetype, filesharing, editing, etc. Clippy is just a meme now and that's a likely future scenario for the "chat" features in LLM products IMHO.
How was the model trained? The gpt-oss models were pretrained primarily using synthetic data along with some heavily filtered real code. The models were then post-trained using distillation (RLKD against o3/hailmary) and berry. For more details about the model training process, please refer to the documentation.
Meaning the author was spot on.
Amazing
Do you care about binary blobs in the kernel? No. Are binary blobs in the kernel open source? No.
But it is tedious to go through the same discussion every 10 years, with a relentless industry that wants to dupe people.
If there wasn't a benefit in for them, they would not call it open source.
This means given everything you have two options: 1) train similar model yourself 2) trust model provider. In software you can get script and run, or get code and compile it in exactly the same binaries.
Naturally 'open source' has different meaning. Some are trying to monopolize it, like they know the 'truth'. Others simply ignore it. Eventually we'll settle on something.
I've never tried self training an AI model or anything but I'd be very curious about the process. Maybe this is something someone could do as a side hustle?
[1] Uncensor any LLM with abliteration by Maxime Labonne https://huggingface.co/blog/mlabonne/abliteration
I think the mention of the "horny people" is warranted, they are an important part of the open models (and first to explore the idea of "identities / personas" for LLMs, AFAIK). Plenty of fine-tuning bits of know-how trickled from there to the "common knowledge".
There's a thing that I would have liked to be explored, perhaps. The idea that companies might actually want what -oss offers. While the local llm communities might want freedom and a horny assistant, businesses absolutely do not want that. And in fact they spend a lot of effort into implementing (sometimes less than ideal) guardrails, to keep the models on track. For very easy usecases like support chatbots and the like, businesses will always prefer something that errs on the side of less than useful but "safe", rather than have the bot start going crazy with sex/slurs/insults/etc.
I do have a problem with this section though:
> Really open weight, not open source, because the weights are freely available but the training data and code is not.
This is factually incorrect. The -oss models are by definition open source. Apache2.0 is open source (I think even the purists agree with this). The requirement of sharing "training data and code" is absolutely not a prerequisite for being open source (and historically it was never required. The craze surrounding LLMs suddenly made this a thing. It's not).
Here's the definition of source in "open source":
> "Source" form shall mean the preferred form for making modifications, including but not limited to software source code, documentation source, and configuration files.
Well, for LLMs the weights are the "preffered form of making modifications". The labs themselves modify models the same as you are allowed to by the license! They might use more advanced tools, or better datasets, but in the end the definition still holds. And you get all the other stuff, like the right to modify, re-release, etc. I'd really wish people would stop proliferating this open weight nonsense.
Models released under open source licenses are open source. gpt-oss, qwens and mistrals (apache2.0), deepseeks(MIT), etc.
Models released under non open source licenses also exist, and they're not open source because the licenses under which they're released aren't. LLamas, gemmas, etc.
When someone joins your data science team your would give them all this code and data. Not just the weights and say - the weights are the source, modify that to improve the model, I look forward to see your MR next week.
EDIT: Heck, sometimes the way to make improvements (modifications) is just to improve the data, and not touch the training code at all. It is often one of the most powerful ways. You still need training code though, and evaluation to measure the impact.
This is similar to the approach taken by some video game studios: release the source code under a permissive license, but not the game assets. Which is better than a proprietary license, but it still presents a hurdle for the final product to be built from source.
The open weights approach is much more user hostile, however. Proprietary game assets can at least be purchased, and the final product can be built. With open weights, this is not possible. Nobody can realistically build the same model or similar models from weights alone. They can use the weights and self-host the prebuilt model, but not create revisions of it, which is the whole point of open source.
Weights are essentially the bytecode of language models. Sure, you can run and modify it with the right tools, but without the tools used to create it in the first place, the project is not much more useful than publishing binaries.
>The “source code” for a work means the preferred form of the work for making modifications
The GPL refers to a form of the artifact being released
IMO they do. If you can't modify it like a core contributor would, then it's not really open source. Traditional open source projects always included development guides, test configurations etc.
> Software like sqlite don't release their full test suite. They paywall the preferred way of making changes, yet they are open source.
That's a matter of opinion. IMO sqlite is not true open source, for precisely this reason.
> The labs themselves modify models the same as you are allowed to by the license
Do the labs do not use source code?
It is a bit like arguing that releasing a binary executable is releasing the source code. One could claim developers modify the binary the same as you are allowed to.
The weights are part of the source code. When running inference on a model you use the architecture, config files and weights together. All of these are released. Weights are nothing but "hardcoded values". The way you reach those values is irrelevant in the license discussion.
Let's take a simple example: I write a chess program that is comprised of a source file with 10 "if" statements, a config file that matches between the variables used in the if statements and a "hardcoded values" file that stores the actual values. It would be a crappy chess program, but I hope you agree that I can release that as open source and no-one would bat an eye. You would also be granted the right to edit those hardcoded values, if you wish so. You'd perhaps make the chess bot better or worse. But you would be allowed to edit it, just like I would. That's the preferred way of modifying it. Me providing the methods that I used to reach those 10 hardcoded values has 0 bearing on my crappy chess bot being open source or not. Do we agree on that?
Now instead of 10 values, make it 100billion. Hey, that's an LLM!
> It is a bit like arguing that releasing a binary executable is releasing the source code.
That's the misconception. Weights are not a binary executable. In other words, there isn't another level above weights that the labs use to "compile" the weights. The weights exist from the beginning to the end, and the labs edit the weights if they want to modify the models. And so can you. There isn't a "compilation" step anywhere in the course of training a model.
The training data is the source code and the training process is the compiler. There's a fairly direct metaphor to be made there. Different compilers can have vastly different impacts on the performance of the compiled executable.
If you will allow me the absurd analogy: my arm is also part of a person (me), but my arm is not a person. My arm does not have its own bank account and pays taxes independently.
I get how the weights are not exactly like binary code. Good points. But they are also not source code (from your own quote)
> "Source" form shall mean the preferred form for making modifications
The weights are not the preferred form of making modifications. At most, one could argue it is the weights + source code.
> In other words, there isn't another level above weights that the labs use to "compile" the weights
The source and training data?
I see your points, and it is an interesting discussion of nuances, but I profoundly disagree that the weights are "the preferred form for making modifications". For these reasons I prefer the term "open weights" for these projects.
I also believe the four freedoms are violated to some extent (at least in spirit) by just releasing the weights and for some that might be enough to call something not open source. Your "freedom to study how the program works, and change it to make it do what you wish" is somewhat infringed by not having the training data. Additionally, gpt-oss added a (admittedly very minimal) usage policy that somewhat infringes on the first freedom, i.e. "the freedom to run the program as you wish, for any purpose".
This isn't entirely about being a beginner or not either. Full fine-tuning without forgetting does really want the training data (or something that is a good replacement). You can do things like LoRa but, depending on your use case, it might not work.
Most "vibes" people are missing that it as only has 5B active parameters.
They read 120B and expect way more performance than a 24B parameter model, even though empricaly a 120B model with 5B active parameters is expected to perform right around there.
Consider the following: it is possible to release binaries under the Apache2 license. Microsoft has, at least at one point, released a binary under the BSD license. These binaries are not open source because they are not source.
This isn't the same argument as given in the article though, so I guess it is a third position.
Agreed. But weights are not binaries in the licensing context. For weights to be binaries it would imply another layer of abstraction, above weights, that the labs use as the preferred way of modifying the model, and then "compile" it into weights. That layer does not exist. When you train a model you start with the weights (randomly initialised, can be 0 can be 1, can be any value, whatever works best). But you start with the weights. And at every step of the training process you modify those weights. Not another layer, not another abstraction. The weights themselves.
If that were relevant to the licensing discussion, then you'd have to consider every "generated" parts (interfaces, dataclasses, etc) of every open source project artefacts. Historically, that was never the case. The license doesn't care if a hardcoded value was written by a person or "tuned" via a process. It's still source code if it's the preferred way of modifying said code. And it is. You can totally edit them by hand. It would not work as well (or at all), but you could do it.
On the other hand, I don't really think that argument then extends to model weights, which are not just some number of steps removed from source code, but just simply not really related to source code.
Of course. This is indeed the fact. /s
See that "/s"? I did write it here, but it's suspiciously absent in the original text. Almost makes you think it was typed in all seriousness.