Artificial Intelligence
- Started 11 years ago
- Last post 3 hours ago
- 1,863 Responses
- renderedred0
These AI models would rather hack than play fair
The researchers tested multiple LLMs, including OpenAI's GPT-4o, Anthropic's Claude 3.5 Sonnet, and DeepSeek R1, to see how they would handle a seemingly straightforward task: playing chess against Stockfish, one of the strongest chess engines in existence
- grafician-2
- This is the problem the democrats created for themselves... they snubbed their nose at everyone that wasn't like them and it cost them the election.canoe
- https://www.porchlig…canoe
- "Majority rule, don't work in mental institutions"BabySnakes
- YakuZoku9
An AI robot got aggressive towards a crowd at an event in China.
- So it begins... Planet of the Robotsyuekit
- Thugbotsdbloc
- Come at me bro!utopian
- aye...renderedred
- It lost its balance?canoe
- I thought that's kanyepango
- splash a glass of water on him. done.Krassy
- LOL @pangoKrassy
- Haha pangoPonyBoy
- Aye mon, thenmaikel
- LOL.
How much are we reading into this?
I just see a robot glitching but the vest really changes the perspective.palimpsest - you fucking what!?milfhunter
- grafician-2
3.7 Sonnet it's here
- https://www.anthropi…grafician
- Is this independent research?
IF it's coming from themselves, it would be pretty biased, no?utopian - Are you okay? Are you a bot?grafician
- used it for a bit, liked chatGPT better for writing dialog oddly as Claude is supposed to be better.YakuZoku
- I thought Claude is supposed to be good for coding.yuekit
- Grok numbers are def fake news misinformation. LOLKrassy
- No, it's actually goodgrafician
- yuekit: Cursor just added the latest version and people say it's the best for nowgrafician
- All AI benchmarks fake alignment, they get 100% at faking it, slightly worse at sthg else, then change the benchmark... I do like Claude tho...kingsteven
- Deepseek has been pretty awful for reasoning, chat gpt skims and forgets everything you just told it (and openai is shit)kingsteven
- renderedred0
The Mask Comes Off: A Trio of Tales
- yuekit0
It feels like LLMs are becoming indistinguishable in their capabilities. ChatGPT, Gemini, Grok, DeepSeek...they can all answer questions, generate copy and solve coding problems well enough for most people most of the time. They are useful tools but I'm not sure I see the path to "AGI" or superintelligence.
- The path might be as simple as the incremental introduction of more and more specialized models that do what people used to. Eventually they can do everything.monNom
- Maybe...but there are still a lot of gaps that make it difficult to replace humans. Right now AI is like a productivity tool for people who already have someyuekit
- technical skills rather than a replacement for someone's entire job.yuekit
- AGI is already here; Superintelligence will be here by end of next year.Krassy
- There is no AGI, it's just hype for nowgrafician
- AGI is already hereKrassy
- There's no universal definition of AGI so sure maybe it's already here...or not.yuekit
- But for me general intelligence would include the ability to learn things from scratch and current AI can't really do that.yuekit
- There is no AI, it's just hype for nowgrafician
- AI is already thereKrassy
- hah, im with graf on this. LLMs are great and AI all smoke and mirrors, anything nearing AGI i'd say definitely at least a decade off.kingsteven
- although maybe the current generation of AI will get us there quicker.kingsteven
- Transformers and LLMs are not a path to AGI or any AI, it's just hype for nowgrafician
- not LLMs but genuinely interesting things happening with AI for example big data LLMs vastly improving scientific researchkingsteven
- *big data AIs, not LLMs durrkingsteven
- mg333
Claude is destroying ChatGPT for me in terms of coding. I started building generative art tools last year and hit limits with GPT in terms of accuracy, quality, and ability to follow instructions.
I was amazed that I could take one of the tools I built that I'd maxed out in GPT (800 lines of JS), drop it into Claude, ask it to tell me what it thought the code did (100% accurate and insightful), and then immediately add new features flawlessly.
Since someone in a Slack group suggested Claude,I haven't been able to stop working on the things I'm building. I can code but not much JS and I'm just astounded at its ability to create code that does what I need it to. After I wrap up these projects I'm moving on to learning more X code to build an app I've had on my mind for a while. I was starting that in GPT but Claude is going to be so much better.
- Word is the old one 3.5 is still good, the new 3.7 it's "too much"grafician
- still not capable of building secure, reliable applications IMO. i still have to get my hands dirty.kingsteven
- I'm expecting that for sure, I just love what quick work it makes of things that would otherwise be an extensive tutorial, or random code snippets.mg33
- totally, as a coder sometimes it seems like i spend more time itemising development to frame it in a way the LLM can give reliable results than it would take tokingsteven
- do it myself, but it's usually worthwhile as i end up with a full featured app in a week rather than a working prototypekingsteven
- I've moved all my tools to my local server, but I'll have to share them at some point. It's basically things that you could do in PS or Illustrator, butmg33
- generatively and quickly. One is a generative collage tool with various parameters, randomized image assets from my assets folders, shapesmg33
- lines, layer effects, etc. The other is a grid based pattern tool with SVG shapes, and ability to adjust padding + and - for overlay effects with blend modes.mg33
- Both create art like this: https://www.instagra…mg33
- It's fun, because it's taking me back to the art I liked making in PS in the early 2000s, but generatively, and with code. It made being out of work last yearmg33
- tolerable because it kept my more engaged than even some work-based project work could.mg33
- sorry, but I can create all that art in any vector app even Figma in like 5 mins man
Whay do you need to code apps to output that?grafician - Instead I would code some apps and put them in the Appstore and get cash for the effort
"Vibe coding" is shitgrafician - Where can I go to learn********
- ask chatgpt :))grafician
- graf - The randomization of it is what's enjoyable. I could make that in Figla, Illustrator, or PS quickly too, but I enjoy building these things and it helpsmg33
- me generate a wide range of ideas quickly that I can then import SVGS from and further customize, incorporate images, etc.mg33
- renderedred2
AI and Esoteric Fascism
- grafician-3
deepseek released their inference code open source
lol profit margin over 500% :))))
- Not to mention they invented their own data center file system for training
https://x.com/deepse…grafician - all their infra code and knowledge
https://github.com/d…grafician - i won't ever use deepseek.plash
- no advanced voice mode, no thanks, typing in prompts is so 2024YakuZoku
- Skills issuegrafician
- Not to mention they invented their own data center file system for training
- renderedred1
There is a Model for That: Science and Public AI Infrastructures
- renderedred0
Roko's basilisk
- renderedred0
Echosent
- webazoot0
'They wanted to save us from a dark AI future. Then six people were killed'
https://www.theguardian.com/glob…
(This is a long and also just plain nuts. Also could've gone in the Conspiracy of the Day thread or the WTF one, did we have a Long Read thread. I couldn't find it. If it looks too long I'm sure it will end up a six part series on Netflix, eventually.)
- *(This is a long and also just plain nuts READ.webazoot
- oh, the zizians,of courserenderedred
- insane bunchrenderedred