Artificial Intelligence thread

[BlackWindMnt]

Not sure if you guys were aware but there are some developer influencers streaming 24/7 one week of game development with mostly claude prompting on twitch and youtube. Just look for theprimetimeagen/theprimeagen on youtube and twitch.
They are all experienced senior developers so this is really interesting view regarding llm (vibe)coding for all to see.

 

[OptimusLion]

Improvements in DeepSeek-V3-0324 (more complete on the official account):

- Reasoning performance: The new V3 model draws on the reinforcement learning technology used in the training process of the DeepSeek-R1 model, which greatly improves the performance level in reasoning tasks, and achieves scores higher than GPT-4.5 on mathematics and code-related evaluation sets.
- Front-end development
- Chinese writing upgrade: The new V3 model is further optimized based on the writing level of R1, and at the same time, the content quality of medium and long text creation is particularly improved.
- Chinese search capability optimization: The new V3 model can output more detailed and accurate content for report generation instructions in online search scenarios, and the layout is clearer and more beautiful.
- Others: tool calls, role-playing, Q&A chat

https://twitter.com/i/web/status/1904526863604883661

 

[AI Scholar]

I don’t believe AI progress is slowing down at all... If you focus solely on China, you’ll see an exceptionally rapid trajectory of advancement. We’ve gone from relatively weak models in 2023 to Qwen2.5 approaching GPT-4’s level last year, and now the latest DeepSeek model became the top non-reasoning AI. What is slowing down is AI progress in the U.S, so I’d argue that China is likely to take the lead by the end of 2025. Once that happens, it won’t look like progress is stagnating anymore, because China will be the one leading all the benchmarks.

 

[Eventine]

Speaking of AI developments, Gemini 2.5 just dropped

Please, Log in or Register to view URLs content!

Claims to have caught up to the state of the art across the board, and beats Claude 3.7 on certain bench marks like AIME and Humanity's Last Exam. Not sure how it compares to the new Deep Seek, but it certainly seems to crush the old R1. Oh, and features the Google classic 1 million context window with a 2 million context window soon to ship. Though of course, we don't know yet how well it deals with those windows beyond just supporting them.

This IS a thinking a model, but if Google keeps with their 1,500 free API requests a day, Open AI and Anthropic are going to be crying to sleep tonight.

As for Deep Seek, looking forward to R2, which based on the performance of V3.1, should be able to beat the new state of the art, but we'll see.

Competition seems to be heating up again.

 

[9dashline]

What? OpenAI o1 Pro api is $600 per million tokens ?!?!!

 

[sndef888]

Honestly, thank god for China

I was scared shitless America was going to dominate the world for the next 100 years again using AI.

The open source movement has totally destroyed the ability of American corporations to control this industry and suck all the world's capital like they did in software in the 2000s.

Now, all we need is for Deepseek to upgrade its servers so it doesn't show server busy half the time lol

 

[AI Scholar]

I tested DeepSeek-V3-0324 on the API, and it appears to be completely uncensored. I’m quite impressed with its overall capabilities, especially for coding and creative writing. DeepSeek-R1 still leads in creative writing on EQ-Bench, so it’ll be very interesting to see how far R2 pushes it too.

 

[iewgnem]

Honestly, thank god for China

I was scared shitless America was going to dominate the world for the next 100 years again using AI.

The open source movement has totally destroyed the ability of American corporations to control this industry and suck all the world's capital like they did in software in the 2000s.

Now, all we need is for Deepseek to upgrade its servers so it doesn't show server busy half the time lol

Considering "American" AI is just Walmart-Chinese AI, and their non-Chinese management's approach has been brute force with hardware, it was never going to work out.

Not to say China will achirve AGI anytime soon either, after all by definition any AGI should be able to at least improve itself and diverge.

But Chinese focus on efficency isnt just about lower cost either, like manufacturing, being about to do the same thing at lower cost also means beibg able to do much better at same cost. If you're able to run a 600B parameter model on $10k hardware, how many parameter and how much reasoning can you do with $1M hardware, or $1B, etc

 

[mossen]

I don’t believe AI progress is slowing down at all...

I don't know how this meme got started. There was some validity for it back in early 2024 when progress seemed to have slowed down. But not now, where progress is literally counted in months, rather than quarters let alone years.

What is slowing down is AI progress in the U.S

No? Google's latest release shows that top US labs still do very fast improvement. Again, this meme of an AI slowdown is really silly and people who perpetuate it mark themselves as unknowledgeable.

 

[AI Scholar]

I would argue that what the U.S is truly exceptional at, and what shows no signs of slowing down, is AI hype. Gemini 2.5 Pro is an incremental technical upgrade over DeepSeek-R1 in many areas, but it’s closed-source, leaving us in the dark about its details. It likely doesn’t surpass what we’ll soon see open-sourced with R2. Can this really be called AI progress when it’s a black box, accessible only through Google’s heavily censored API, with no accompanying research paper? Google can’t achieve true AI progress like DeepSeek because its core business model holds it back.

Meanwhile, high-quality, practical AI advancements are flourishing in China, going far beyond chatbots into manufacturing, healthcare, education, electronics, EVs, and many other fields. It’s being integrated everywhere due to DeepSeek’s open-source nature, which allows anyone to host and modify it to their needs.

In my view, meaningful U.S AI progress has undoubtedly slowed over the past six months (since the reasoning paradigm was proven by o1), while the hype around U.S AI only grows louder. In contrast, China’s meaningful AI progress has accelerated significantly since the release of Qwen-2.5 six months ago, followed by DeepSeek V3/R1 and the imminent R2. Add to that BYD’s integration of AI into all its cars, the rapid advancements by countless humanoid robot companies, and the profound economic ripple effect created by R1 across China’s entire economy. This is real AI progress. Not Gemini’s censored, closed black box, which will become irrelevant the moment R2 arrives.