Artificial Intelligence thread

[meedicx]

https://twitter.com/i/web/status/2062389885437366342

The narrative has definitely shifted this week with US mainstream media now aware of what was being said in this thread; lots of attention on token cost optimization and switching to cheaper models.

Some other examples:

https://twitter.com/i/web/status/2061862733126275549

AI harness startup (Droid) announces a smart model routing product to reduce costs

https://twitter.com/i/web/status/2062218656420167785

Legal AI startup Harvey saw significant cost reductions routing most tasks to GLM 5.1. Was also able to surpass frontier models by fine tuning GLM directly, reinforcing my previous point that fine-tuned open weight models will surpass US closed models for domain specific cases.

https://twitter.com/i/web/status/2062900125169537436

Former OpenAI expert interview says many of his clients have switched to open weight models. Chinese model released in April (ie. DeepSeek v4 and Kimi K2.6) have seen vast improvements in quality making them very competitive.

 

[iewgnem]

Using Claude to manage inboxes, meetings, and calendars is completely pointless. I'm surprised they've managed to stay in business and are only now switching to Deepseek.

They were most likely using Haiku or Sonnet without thinking before; their quality, speed, and price were absolutely crushed by V4 Flash.

I'm still a big fan of Anthropic and think they'll be fine because of enterprise relationships + dev brand + eventually hopefully ramping up on capacity + their next generation of models + moving up the stack. But the Chinese models really are on their heels and I assume will apply significant margin pressure on API for a long time.

As I said, Anthropic's entire business model is scaming enterprise management.
And I know from personal experience a lot of them can't even tell the difference between Opus and Sonnet.

At end of the day the cold hard financial reality doesn't care how much branding you have, you either switch to Chinese models, or you just pick how you want to die.

 

[bsdnf]

The narrative has definitely shifted this week with US mainstream media now aware of what was being said in this thread; lots of attention on token cost optimization and switching to cheaper models.

Some other examples:

https://twitter.com/i/web/status/2061862733126275549

AI harness startup (Droid) announces a smart model routing product to reduce costs

https://twitter.com/i/web/status/2062218656420167785

Legal AI startup Harvey saw significant cost reductions routing most tasks to GLM 5.1. Was also able to surpass frontier models by fine tuning GLM directly, reinforcing my previous point that fine-tuned open weight models will surpass US closed models for domain specific cases.

https://twitter.com/i/web/status/2062900125169537436

Former OpenAI expert interview says many of his clients have switched to open weight models. Chinese model released in April (ie. DeepSeek v4 and Kimi K2.6) have seen vast improvements in quality making them very competitive.

I believe that Cursor’s training of Composer2 using Kimi 2.6 already demonstrates the power of open-source models: LLM companies typically bring general-purpose models to market rather than specialized ones, leaving ample room for post-training and adjustment.

With the depletion of internet data and the ever-increasing prices of existing models, private databases and business-oriented fine-tuning will be the next goldmine. Cybersecurity-specific versions released by Claude and Gpt are one of the signs.