ShockedPikachu.svg
I just tried it on deepseek it did it fine and gave the source for everything it mentioned as well.
But every techbro on the planet told me it’s exactly what LLMs are good at. What the hell!? /s
Not only techbros though. Most of my friends are not into computers but they all think AI is magical and will change the whole world for the better. I always ask “how can a blackbox that throws up random crap and runs on the computers of big companies out of the country would change anything?” They don’t know what to say but they still believe something will happen and a program can magically become sentient. Sometimes they can be fucking dumb but I still love them.
BBC is probably salty the AI is able to insert the word Israel alongside a negative term in the headline
News station finds that AI is unable to perform the job of a news station
🤔
But AI is the wave of the future! The hot, NEW thing that everyone wants! ** furious jerking off motion **
I learned that AI chat bots aren’t necessarily trustworthy in everything. In fact, if you aren’t taking their shit with a grain of salt, you’re doing something very wrong.
This is my personal take. As long as you’re careful and thoughtful whenever using them, they can be extremely useful.
Neither are my parents
But the BBC is increasingly unable to accurately report the news, so this finding is no real surprise.
Turns out, spitting out words when you don’t know what anything means or what “means” means is bad, mmmmkay.
It got journalists who were relevant experts in the subject of the article to rate the quality of answers from the AI assistants.
It found 51% of all AI answers to questions about the news were judged to have significant issues of some form.
Additionally, 19% of AI answers which cited BBC content introduced factual errors, such as incorrect factual statements, numbers and dates.
Introduced factual errors
Yeah that’s . . . that’s bad. As in, not good. As in - it will never be good. With a lot of work and grinding it might be “okay enough” for some tasks some day. That’ll be another 200 Billion please.
I’ll be here begging for a miserable 1 million to invest in some freaking trains and bicycle paths. Thanks.
Is it worse than the current system of editors making shitty click bait titles?
Surprisingly, yes
alternatively: 49% had no significant issues and 81% had no factual errors, it’s not perfect but it’s cheap quick and easy.
Flip a coin every time you read an article whether you get quick and easy significant issues
It’s easy, it’s quick, and it’s free: pouring river water in your socks.
Fortunately, there are other possible criteria.
Idk guys. I think the headline is misleading. I had an AI chatbot summarize the article and it says AI chatbots are really, really good at summarizing articles. In fact it pinky promised.
As always, never rely on llms for anything factual. They’re only good with things which have a massive acceptance for error, such as entertainment (eg rpgs)
I tried using it to spit ball ideas for my DMing. I was running a campaign set in a real life location known for a specific thing. Even if I told it to not include that thing, it would still shoe horn it in random spots. It quickly became absolutely useless once I didn’t need that thing included
Sorry for being vague, I just didn’t want to post my home town on here
You can say Space Needle. We get it.
The issue for RPGs is that they have such “small” context windows, and a big point of RPGs is that anything could be important, investigated, or just come up later
Although, similar to how deepseek uses two stages (“how would you solve this problem”, then “solve this problem following this train of thought”), you could have an input of recent conversations and a private/unseen “notebook” which is modified/appended to based on recent events, but that would need a whole new model to be done properly which likely wouldn’t be profitable short term, although I imagine the same infrastructure could be used for any LLM usage where fine details over a long period are more important than specific wording, including factual things
The problem is that the “train of the thought” is also hallucinations. It might make the model better with more compute but it’s diminishing rewards.
Rpg can use the llms because they’re not critical. If the llm spews out nonsense you don’t like, you just ask to redo, because it’s all subjective.
Or at least as an assistant on a field your an expert in. Love using it for boilerplate at work (tech).
What temperature and sampling settings? Which models?
I’ve noticed that the AI giants seem to be encouraging “AI ignorance,” as they just want you to use their stupid subscription app without questioning it, instead of understanding how the tools works under the hood. They also default to bad, cheap models.
I find my local thinking models (FuseAI, Arcee, or Deepseek 32B 5bpw at the moment) are quite good at summarization at a low temperature, which is not what these UIs default to, and I get to use better sampling algorithms than any of the corporate APis. Same with “affordable” flagship API models (like base Deepseek, not R1). But small Gemini/OpenAI API models are crap, especially with default sampling, and Gemini 2.0 in particular seems to have regressed.
My point is that LLMs as locally hosted tools you understand the mechanics/limitations of are neat, but how corporations present them as magic cloud oracles is like everything wrong with tech enshittification and crypto-bro type hype in one package.
They were actually really vague about the details. The paper itself says they used GPT-4o for ChatGPT, but apparently they didnt even note what versions of the other models were used.
I’ve found Gemini overwhelmingly terrible at pretty much everything, it responds more like a 7b model running on a home pc or a model from two years ago than a medium commercial model in how it completely ignores what you ask it and just latches on to keywords… It’s almost like they’ve played with their tokenisation or trained it exclusively for providing tech support where it links you to an irrelevant article or something
Gemini 1.5 used to be the best long context model around, by far.
Gemini Flash Thinking from earlier this year was very good for its speed/price, but it regressed a ton.
Gemini 1.5 Pro is literally better than the new 2.0 Pro in some of my tests, especially long-context ones. I dunno what happened there, but yes, they probably overtuned it or something.
Bing/chatgpt is just as bad. It loves to tell you it’s doing something and then just ignores you completely.
I don’t think giving the temperature knob to end users is the answer.
Turning it to max for max correctness and low creativity won’t work in an intuitive way.
Sure, turning it down from the balanced middle value will make it more “creative” and unexpected, and this is useful for idea generation, etc. But a knob that goes from “good” to “sort of off the rails, but in a good way” isn’t a great user experience for most people.
Most people understand this stuff as intended to be intelligent. Correct. Etc. Or they At least understand that’s the goal. Once you give them a knob to adjust the “intelligence level,” you’ll have more pushback on these things not meeting their goals. “I clearly had it in factual/correct/intelligent mode. Not creativity mode. I don’t understand why it left out these facts and invented a back story to this small thing mentioned…”
Not everyone is an engineer. Temp is an obtuse thing.
But you do have a point about presenting these as cloud genies that will do spectacular things for you. This is not a great way to be executing this as a product.
I loathe how these things are advertised by Apple, Google and Microsoft.
-
Temperature isn’t even “creativity” per say, it’s more a band-aid to patch looping and dryness in long responses.
-
Lower temperature is much better with modern sampling algorithms, E.G., MinP, DRY, maybe dynamic temperature like mirostat and such. Ideally, structure output, too. Unfortunately, corporate APIs usually don’t offer this.
-
It can be mitigated with finetuning against looping/repetition/slop, but most models are the opposite, massively overtuning on their own output which “inbreeds” the model.
-
And yes, domain specific queries are best. Basically the user needs separate prompt boxes for coding, summaries, creative suggestions and such each with their own tuned settings (and ideally tuned models). You are right, this is a much better idea than offering a temperature knob to the user, but… most UIs don’t even do this for some reason?
What I am getting at is this is not a problem companies seem interested in solving.They want to treat the users as idiots without the attention span to even categorize their question.
-
Which is hilarious, because most of the shit out there today seems to be written by them.
Fuckin news!
