🏳️⚧️ Trans | 🏳️🌈 Pan | 💻 Tech | 🎮 Game Dev JavaScript, Node, Express, C#, Unity, Godot, AI
108 followers 295 following 273 posts
view profile on Bluesky
No, the thing is the people telling you to kill yourself. It isn't uncharitable, it is bordering criminal, especially if you went and did it. We're not going to just breeze past suicide encouragement when my entire point was that. It is ok for you to not like AI, it is not ok to attack people.
Not liking it is fine, attacking people who do, is not. I've been told to kill myself by anti-AI people because I had the audacity to suggest both people should be respected.
You have a right to control your data and identity. You have no idea what the service he's using is doing with your voice, your image if video, or anything else. If they asked me, I'd tell them sure. If they didn't ask, I'd tell them we'll be going with another contractor if you pull that again.
Not in a home contractor environment, but interviews, business calls, client support calls I see it quite regularly. I don't view it as invasive if they ask, but not asking is incredibly invasive. I like AI, but consent is vital in all of our lives, now more than ever.
Announced today, the initial results of GPT5 for hallucinations is a massive improvement: x.com/polynoamial/...
Most preprints get published to arXiv before being published to a journal. Their paper and dataset and contact info are all available if you'd like to lead that charge.
The paper was published as a preprint in mid April along with the dataset used for anyone to review, it has barely existed long enough to be peer reviewed. The study was done with and by PhD virologists in the field and their contact info is available if you know other expert virologist for review.
[2504.16137] Virology Capabilities Test (VCT): A Multimodal Virology Q&A Benchmark share.google/anFYo1CH01Lh...
The tests were done in collaboration with virologists. Virology Capabilities Test share.google/7gyljUuOHRjU...
It was evidence that 17% may not be close to accurate and both results should be questioned. Competition leaderboards show evidence that it doesn't hallucinate nearly that much. So we should verify what is tested in both the leaderboard and the original chart, what is repeatable and what isn't.
I think that sounds like a great thing to review when it releases. Current models are already outperforming PhDs in the lab, so I think it is important to focus on eliminating hallucinations more than increasing knowledge at this point.
You're referring to a hallucination rate of ~4% as a problem when a person is more likely to lie to you just because they're uncomfortable or nervous. I agree that improvement to as close to 100% should always be the goal but at 4% you're more likely to get a lie from a person.
They're not, they're being mixed together. You can toggle reasoning on or off in the latest OSS models, and I imagine the next big release from OpenAI will be the same.
So it hallucinates significantly less than people lie and the original chart shows 3.7 being 5 times worse than it actually is. 0.1% is within the statistical margin of error, they're likely the same but with Claude 4 being even more capable.
Then it doesn't matter what I say and you're just wasting time denying that the latest models hallucinate less which is shown in the graph. Reasoning models do hallucinate more and the latest non-reasoning models hallucinate very little, even less than shown in the graph.
No, you're talking about two different categories of models, which I've already addressed, nothing was ignored on my end at all. 🤷♀️ You, however, have been completely silent regarding any of that or the fact even newer models exist now with drastically lower hallucination rates reported.
But it does, unless you just ignore the latest models are the two with the least hallucinations.
It doesn't follow that new models will be worse, that's the opposite of logical, but we will have to wait and see. It definitely hasn't been the case for the regular models.
Yeah, we were discussing hallucinations and how the chart showed the latest models had the least hallucinations. Newer models, like Claude 4 Sonnet are between 2-4% hallucination rates and that's the smaller model. We're waiting for the new GPT models now.
The latest models from Anthropic, Claude 4 Sonnet are currently showing hallucination rates of less than 2% but we'll know more as more tests are done.
Grok is from Musk, a narcissist and liar who literally trained the model to refuse to create negative content about him or Trump, I'm not surprised it lies in the slightest. I've never used it. Claude is the most trustworthy, built for research and science. ChatGPT is second, built for consumers.
o-Series models are a different model type altogether and were focused on reasoning, which allowed the models to solve much more complex problems but results in more hallucinations since the model is "thinking" and hypothesizing based on whatever data it has available.
No, GPT 4.5 is their biggest and most powerful general model, it is the largest generational model at 17 trillion parameters. However, it is slow and incredibly expensive. 4.1 was released as a step back to refocus on efficiency at 1.7 trillion, this did not perform well as you can see.
I'm more curious if the trend holds or if GPT 5 will hallucinate more or less. There is a problem with model collapse and diminishing returns on larger models. So I'd be interested in seeing if OpenAI can continue a downward trend in hallucinations or if GPT 5 is wholly unreliable.
Yes, the chart is sorted low to high, but it also shows newest models hallucinated less. You can extrapolate that from the data on the chart and model generations.
GPT 4.5 and Claude 3.7 Sonnet were the latest models from each. As for o3 and o1, they were released a month apart. So in general yeah, they're the same generation. General availability of o3 was 3 months later, the numbers aren't representative of generations with OpenAI. Grok is the outlier.
What I hate so much is the big model censorship. Claude Sonnet is absolutely incredible at roleplay and narrative generation in dynamic RPG games. It rarely outright refuses to perform, but it is hard to risk an account ban if they somehow think your story about a knight murdering someone is real.
I would recommend trying Hugging Faces Spaces to trial the model you're interested in, see if it justifies the cost for your use case. I've found that anything above 70B has diminishing returns when compared to fine tuning and specialization. Leave the big general models to corps if unnecessary.
Yep, and I think that is why AI is beginning to excel in hard science. Things that are quantifiably unbiased like physics and math. We may find new algorithms, new formulae, new physical properties, but the underlying mechanics are inherently solid (for now lol) and AI adapts well to that.
I dislike copilot outside of Github. I prefer Cline with Gemini 2.5 pro planning and Claude 4 Sonnet coding if I had to use big models. Then I run PR review with GitHub's Copilot and my own eyes. But I'm super interested in where Qwen goes next, 32B is scoring in line with Claude 3.5 Sonnet.
I think the new Qwen 3 models are looking good. The real issue is hosting the bigger ones, but even the 32B model is scoring quite well and fits on a 5090. I'd love to see GPUs with high vram for AI that aren't $1-20,000 for no reason. I'd spend $3-$4000 for a 5090 with 160gb of vram.
Yeah, it definitely reflects the inherent bias in the data. Worse, there is no real way to eliminate the bias. The data it trains on is human, which is biased depending on the human. Even the reinforcement training is inherently biased by the human annotator. All data is tainted somewhere.
No, it's not remotely AI nihilism. I like AI. I'm not being a nihilist, but y'all are very much pessimists. Saying you aren't going to convince people who like AI by treating them like shit, isn't nihilism, it is reality.
I'm not assuming anything, I said if the pattern holds, the next model this month will be at parity. That isn't an assumption, you are making a lot of assumptions though.
You're right, humans kill each other far more than AI. It isn't comparable, on the whole, humans are trying to eradicate brown people, trans people, and every Palestinian in Gaza But get mad at me because I had the audacity to compare statistics of humans and AI.
I know, politicians are the worst. They'll bomb people for oil and opium and starve people for religious politics. As for me, I value human life, but make up whatever you want.
One set of models is twice as bad, we'll see where the next model lands. 7% of all human communication are lies, that's all communication not an average of human tendency to lie. The comparison was the percentage of total communication to basically what is being looked at as total AI output.
This is why you'll never convince the billion people using AI every week. You can't have a logical discussion, I'm either stupid or obtuse and you're the only one who knows how things work. What's next? You going to tell me to kill myself like the last anti-AI rager who couldn't control their anger
Twice as unreliable can be not far off, do you consider 2 far from one? You're applying zero logic to your reasoning, this is like Nvidia saying they're twice as fast as previous cards and showing percentages with no actual numbers because the data was twice as fast with fake frames. Data matters.
You have zero data to even find the most honest human or any evidence to point to them being statistically more honest than the average. We do have data that says 7% of total human communication are lies.
Then how do you justify that argument when it becomes more reliable? So far AI models have improved by half each generation, meaning if that standard holds, the next generation model launching some time this month will be as reliable or more reliable than a human. Where do you move the goal post?
Thanks for the info! I believe game makers mp_grid is an implementation of A* the description of the system is very similar. Your game is coming along great!
As for embracing the technology, you're ignoring that each major model has improved on this metric by half. If that holds for the next generation of models coming this month, that would make them either as likely or less likely to lie than a human. See how the argument falls apart?
And? You've shared nothing about how you tested, what the metric was for success, what the tolerance was for error, or even what you were promoting. You are just expecting me to run with whatever you say and that in itself is a big reason why people fail at using AI. No worries, I'm going to work
But you are talking about them as if they are, which would be intentionally misleading. You shared nothing about your use case except implying that you expected the result to be deterministic. So at this point, you're not behaving much better than what you claim AI to be doing.
It is only a lot worse depending on the tolerance to error. Twice as bad doesn't mean bad. If the tolerance for error is 30% it is still far beyond acceptable. If the tolerance of poop shoveling is 30% and AI had an error rate of 15% and can self correct, why would you force a human to shovel shit?
7% of total communication is 7% of total communication, that is in fact the average across all forms of communication (it is). If we break down specific areas of communication that number can be significantly higher or lower.
If you'd like to discuss that, we should move somewhere without a character limit as this isn't conducive to a real discussion.
That boils down to you not understanding what you're using. It didn't used to be better, it quite literally used to be horrible. Newer models are significantly better in general. As for why it generated different results, LLMs are not deterministic in nature.
I completely disagree. Early models, as shown in the chart, were statistically and quantifiably much worse. I'm glad you're experimenting with AI, but you clearly want to blame it for its failure rather than take equal ownership of your partnership in that failure. AI isn't always right.
Another possibility is that you're just not good at using AI, it isn't generally intelligent and how you prompt it and what context it is given matters a lot. The same goes for a human, what you ask, how you ask and what data they have greatly matters. People will pretend to know things they don't.
Considering various models have allowed me to build entire applications and research topics with cited materials I can verify, I have experience of it being fairly great but ok. If its worst day is 15, it's not far off from a person and you must think people are kinda terrible instead of fairly so.
So then do you feel people already lie a lot? Because yes, it's only twice that of a human, which is about 7% of total communication is a lie.
No you said on its best day, these are all different models. You could have just Googled how often people lie and found it, I'm not trying to prove anything to you. The OP cited their statement and it was still intentionally misleading because none of you read and understood the cited data.
That isn't the discussion now is it?
They don't, but in general they do believe they're white lies.
I do, and I also consider the environmental cost of humans has been catastrophic as well. I think that running an AI model locally on my computer is definitely not worse than that of a person.
No, the chart shows the newer models on the left and the older models on the right. Imagine that, the person who posted the chart was being intentionally misleading. Also known as lying.
It generally is better than most people when given the right context.
7% of total communication means in general yes.
I guess that's why so many people get fired for being bad at their job then.
I'm not arguing anything, I'm pointing out that someone is claiming it's completely unhelpful and a waste but they're being completely dishonest as it's not far off from a human.
So it lies only about twice that of humans. Research has shown that about 7% of all human communication is a lie. Or do you believe humans never lie? Even if they didn't lie at all, how often do you think they're confidently wrong? Or cause accidents due to mistakes?
No, they're absolutely right. Most people don't know how to properly use AI and working as a prompt engineer has nothing to do with it. People who know how to use AI are making medical discoveries and finding new ways to treat diseases. People who don't, are asking AI about strawberries.
The 65% margin of error is actually because they just showed a chart of older models next to newer models. Which is funny, because the chart shows AI has improved drastically in not hallucinating with each new model, probably not their intention The left is the most recent models.
Not sure if talking about humans or AI here, could go either way.
If they'd leave, stop giving them tax havens to run to? Immediately trying to gaslight saying NYC is bustling and people are happy with their lot while complaining about the guy those people are voting to make mayor because they're not happy with being constantly cheated by the rich.
Well, I'm not a millionaire and if I was making a movie about an evil AI taking over the world with no budget, I'm definitely using genAI to create and play the AI. If artists want to volunteer for rev share, they're welcome to do that but I'd still use AI for the AI... I'm not Disney.
I just feel like an AI character in Tron of all things makes sense and this was not the right call. Like, if they remake Space Odyssey I would not complain if HAL was an AI, I think it's an appropriate use of the tool to have AI be an AI.
Bought, thanks for the tunes stranger. I hope you're well!
He literally made jokes to him about their "secret" and said he loves young girls, just like him. And people are still like, "they were just friends! Yeah, sure.
Even weirder that it happened in an unusual place and someone high profile was a victim caught in the cross fire. Almost like this was all politically motivated.
Lmao no it wouldn't, they would shout woke and claim you were indoctrinating them. They think science is a ploy by liberals to trans their children. It's not an education problem, it's a hate problem.
Eyyy, King Missle! I have the greatest kept secret in the world if you like John.
You can't reconcile with the revelation that one half of the country wishes the other half was dead or enslaved and elected a felon pedophile rapist sex criminal to lead them in that mission because they LIKE him and couldn't wait to be their true selves openly.
I think the hardest part is realizing there is no peaceful way to repair the divide and watching people yell "Won't someone do something!" and basically mean, "tell them they're mean and they'll stop." referring to people who want them dead or suffering...
This is so awesome, do you ever do commission work for indies? Or just your own thing?
30% reduction, not 70% 🤣
I recently used Cline to refactor the GameSession and Ambiance/Audio Manager and just that was about 3500 lines. Refactor probably cut that in half, maybe 70% reduction. I'll message you.
Bahaha good, then this would be perfect. I have a private repo I've used to build a really complex AI powered TTRPG game for Vampire The Masquerade. I specifically set the rule as LLM code only. How big do you want it? I can add you on signal.
Do you know React, electron, and tailwind? If so, I have something you would probably loooove.
What do you mean by handling large scale traffic? Bluesky is the backbone, no infra needed for a client app that just interfaces with an external backend that's already at scale. I think you confused a client app with a full client/server Bluesky replacement. Each client is 1:1 input to output.
My phone deleted usability, but you probably got the idea.
A weekend for a decent MVP with some bug fixing and maybe some of you have some friends to poke at it nearby. That's probably what Bluesky itself took and then years of development lol
Nice day for fishin, ain't it?
Depends on the genres you like. Metal is definitely Ashes of the Wake (20th Anniversary Edition) by Lamb of God
My general experience is that it is probably closer to 50/50. I let other family members try it as well and my dad and my partner plus my uncle are the only ones I've met who experienced sickness. My aunt and my cousin did not. So that's 3/8 from my limited data pool. My kids are in VR 2hrs a day.
They didn't, but high refresh rate and a wide fov helps. Also not everyone experiences VR sickness. I don't actually experience the issue with VR sickness and none of my three kids do either. We could play all day and I use smooth motion too. My dad and my partner, 30min tops and they're sick.
Oh yeah, there was. It was discontinued, not sure why, it was awesome. A mod adds it now, but like it started up and you were on a couch in a room looking at a TV all in Minecraft and then hopped into a world and it was Minecraft VR.
Or that the next generation after them grew up inside VR Minecraft.
Not only that, but they're drawing down on you and you've even read an article from them. They never had a free article to begin with 🤣
I feel this, Maya was the first time it all just clicked. 3DS Max felt like I was creating with AutoCAD and Blender felt like an alpha product I couldn't use as well. Maya though, I made a whole ship and an animation in like 3 days my first time following a tutorial: youtu.be/F89QfpFhCx8?...
And yes, I know I'm a monster for refusing to ever use a linter or prettier on my code 💀
See, this is why I just stick to my code looking like shit if I can't personally write it to spec 🤣🤣
While testing the new dialogue box I think I've settled on VN style portraits with a name tag, reserving the box itself for text. I feel like VN portraits feel more personable for an RPG. @streetsofblood.bsky.social #gamedev #indiedev
I'm glad you're still here.
It amazes me that this is what gets banned for community guidelines there but they allow MAGA to literally tell people they're going to die when they take over.
This and the dozen comical puns about the horribly named evil villains destroying our society is all the more evidence that we're living in a simulation and the game master has grown tired of our inability to figure it out, so he has played his final Trump card...
Started testing on the dialogue panel and a new apartment scene for @streetsofblood.bsky.social #gamedev #indiedev
It's the exact same attack they made in Patreon and then Subscribestar. They'll be forced to remove anything related to whatever the processor deems as amoral or grotesque, adult content is just the most obvious. Lesser obvious example, Trevor and the teddy bear boyfriend in GTA V is high risk.