Artificial Intelligence thread

[proelite]

My point is that now that agents allow developers to complete complex tasks, then that means you need fewer human software devs. So who is losing jobs?

For senior developers and up, coding is actually half if not less of the responsibilities. So at least for experienced folks, it gets software delivery cadence back to where it was in the 2000s and early 2010s, because software has become much more complex and obviously just throwing more programmers at the problem only made it worse. AI makes it possible to scale productivity in software again.

 

[bsdnf]

My point is that now that agents allow developers to complete complex tasks, then that means you need fewer human software devs. So who is losing jobs?

I know for my firm, I am not hiring anyone right now because myself + AI is more productive than myself with several mid tier devs because I can explain what needs to be done to AI better than another dev & AI can get the code done better than mid tier devs.

it frankly doesn’t matter that opus is better than GLM because GLM can answer all the tasks I need it to do on 1 pass. So when all the LLMs can do the task, I will pick the lowest cost one.

Junior coders are completely F@cked, job prospects are terrible.

Their options are to make demos using GLM/Minimax/Kimi, start their own businesses or at least prove their ability to uncover new demands and products, rather than focusing on coding

Oh, that explains why the use of these models has skyrocketed.

 

[tphuang]

https://twitter.com/i/web/status/2023469180268519828

if you watch openrouter now, the token generation has sky rocketed 2x in the past month.

Chinese models now account for probably half of the token generation or more by this point.

 

[tokenanalyst]

I mean in terms of product, they compete better against GLM and Minimax.

Doubao is on a different level than them.

When I used Qwen 3 web version a few times, the results were always garbage. Which I never had with Kimi 2.5. K2.5 is always excellent. It's puzzling why Qwen-3 large models suck so much.

I think with this version the Qwen team is cooking something good. My biggest problem with only text LLMs is that sometimes describing a problem with words is difficult, the may have the ability to debug code but once the code is compiled free of errors is very difficult for the model to solve a problem and improve the code.

Qwen 3.5 vision capabilities are excellent, it can detect even subtle details in videos and images. Is not as good as other Chinese models at one shot generation but its agentic capabilities looks pretty good. Even if the code doesn't look good in the first attempt it have the ability to improve without degrading too much. I think combining the vision capabilities with its improving agentic capabilities, there is room something really good.

 

[siegecrossbow]

Junior coders are completely F@cked, job prospects are terrible.

Their options are to make demos using GLM/Minimax/Kimi, start their own businesses or at least prove their ability to uncover new demands and products, rather than focusing on coding

Oh, that explains why the use of these models has skyrocketed.

I’m concerned that once seniors retire enmass there won’t be enough juniors to replace them. Can’t wait till management or CEOs start vibe coding non-scalable products because they can’t find anyone to do or maintain it lol.

 

[proelite]

I’m concerned that once seniors retire enmass there won’t be enough juniors to replace them. Can’t wait till management or CEOs start vibe coding non-scalable products because they can’t find anyone to do or maintain it lol.

That’s why we won’t retired. Hahahahaha

 

[bsdnf]

According to the Science and Technology Board Daily report on February 17, just over a month after completing its previous $500 million financing round, Moonshot is about to finalize a new funding round exceeding $700 million. This round is co-led by existing shareholders including Alibaba, Five Yards, and Jiuan, with Tencent also participating.

Additionally, the company has already initiated another funding round at a valuation of $10-12 billion.

 

[tphuang]

View attachment 169833
According to the Science and Technology Board Daily report on February 17, just over a month after completing its previous $500 million financing round, Moonshot is about to finalize a new funding round exceeding $700 million. This round is co-led by existing shareholders including Alibaba, Five Yards, and Jiuan, with Tencent also participating.

Additionally, the company has already initiated another funding round at a valuation of $10-12 billion.

Yes good for moonshot because they didn’t IPO like zai and Minimax so need this private funding round to keep up on cash side of things.

 

[mossen]

Gemini 3.1 Pro-Preview is out. It scores #1 on most benchmarks. But that's not what I'm interested in. The big improvement is the fall of hallucination rates.




My view is that low hallucination rates is perhaps the most important metric for AI progress, because it enables all other metrics. What good is huge intelligence gains without coupled falls in hallucination?

Google does seem like the most likely candidate to win the AI race in the West. The only area where they are behind is in coding, but that's mostly because Demis has a broader view of AI progress. He is more focused on world models and embodied AI, which I think is the smarter long-term play. The coding gains will come with time anyway.

 

[bsdnf]

Gemini 3.1 Pro-Preview is out. It scores #1 on most benchmarks. But that's not what I'm interested in. The big improvement is the fall of hallucination rates.

View attachment 169971


My view is that low hallucination rates is perhaps the most important metric for AI progress, because it enables all other metrics. What good is huge intelligence gains without coupled falls in hallucination?

Google does seem like the most likely candidate to win the AI race in the West. The only area where they are behind is in coding, but that's mostly because Demis has a broader view of AI progress. He is more focused on world models and embodied AI, which I think is the smarter long-term play. The coding gains will come with time anyway.

context window doubled, but still pitifully small