Maybe this is the most important, it’s opensourced
Eventually all these puzzle pieces come together.
Llms are just a 
part of AIs, hardware will also significantly contribute to its collective intelligence…
1 Like
IMO, fear of being replaced and having no future is a legit fear. We don’t know the future, but I certainly feel things are not looking good for future generations of humans as we exist now (sans cyborging or radical genetic engineering).
I’m not going to sugarcoat my posts here - there are great and interesting things about AI, but there is a serious issue regarding replacement of much of humanity and we don’t know how this is going to play out, but it’s dark for sure.
2 Likes
If the memory issue can be overcome, photonic computing will overtake the entire gpu/npu industry.
2 Likes
IMO, this new Mythos model by Anthropic is really good for all of us - in addition to unknown exploits, it will also find hidden back doors that corporations and governments have put into software and possibly hardware too. This is going to force them to seal them up - or otherwise they will be attack vectors for anyone with access to this model or similar models in the future, making the product useless. In fact this is really an amazing advance for all of humanity - leveling the playing field for all.
1 Like
Opus 4.7 released, not to be confused with mythos which I suppose has no number yet.
1 Like
1 Like
Major improvements in image generation.
1 Like
Autogenesis is a self-evolution protocol and runtime for LLM-based agent systems.
Recent agent protocols often under-specify cross-entity lifecycle/context management, version tracking, and safe evolution update interfaces, which encourages monolithic compositions and brittle glue code. Autogenesis addresses this by decoupling what evolves from how evolution occurs:
- RSPL (Resource Substrate Protocol Layer): models prompts, agents, tools, environments, and memory as protocol-registered resources with explicit state, lifecycle, and versioned interfaces.
- SEPL (Self Evolution Protocol Layer): specifies a closed-loop operator interface to propose, assess, and commit improvements with auditable lineage and rollback.
Built on Autogenesis, the system includes an Autogenesis-Agent style tool-calling agent that can dynamically instantiate/retrieve/refine resources and improve during execution.
Kimi K2.6 is now out. It’s about 30-40% more expensive (can still use 2.5 though)
Reddit - Please wait for verification.
Also I see that GPT5.5 has also been released today. I don’t really care about that one as way too expensive for me.
I’ve been exclusively using kimi 2.5 for the last three months and am mostly happy with it. Have been wanting to try MiniMax 2.7 and MiMo 2.5, but currently I am using the kimi-code client, so I am locked into Kimi models with that. I am building my own coding harness however that will allow for multiple providers, so will test with them in the future.
Half a day later and bam!
The details make it look practically impossible. I don’t know how they’ve done this. More to come soon …
New chart with deepseek v4:
I was a bit surprised by this - expected v4 to be more bang for buck than kimi 2.6 … but thinking about it more I’m guessing that kimi has been trained on code for several months now (via k2.5) and I am now using k2.6 and can say that it is a big leap. Many have been using k2.5 for coding and Kimi has definitely been using all of this data for training k2.6. Meanwhile deepseek 3.2 hasn’t been used for coding nearly as much, so less training data for v4.
But Kimi is a 1Terabyte model while the new full deepseek v4 is a 1.7Terabyte model. So I expect that as deepseek gets more training data, v4.x in the future should be able to outpace kimi’s k2.x models. Time will tell.
2 Likes
Surprisingly, V4 Flash can be run with a much smaller footprint of memory. This means local and decent quality inference.
And in V4 cloud news:
X user thehype pointed out that the Chinese AI lab’s discount “is starting a price war in the AI market,” adding:
they just slashed input cache prices to 1/10th of what they already were.
and there’s a separate 75% off promo on v4-pro running until may 5th.
but even ignoring the sales – the normal api prices tell the story. output per 1M tokens (real weighted avg, no discounts):
- gpt-5.5: $30.21
- claude opus 4.7: $25.00
- deepseek v4-pro: $1.73
that’s ~17x cheaper than gpt-5.5 and ~14x cheaper than opus 4.7.
now add the 75% promo: deepseek output drops to $0.87/M. that’s 35x cheaper than gpt-5.5 and 29x cheaper than opus 4.7.
and the benchmarks? v4-pro isn’t that far behind. artificial analysis intelligence index:
- gpt-5.5: 60
- claude opus 4.7: 57
- deepseek v4-pro: 52
13% lower score. 35x lower price.
after releasing v4 on open weights (mit license, free to self-host), deepseek is now aggressively competing on cloud api pricing too. own both ends of the market.
it’s a dangerous game. when a model is 87% as capable at 6% of the cost, “we’re better” stops being a pitch
ai is starting to commodify. the price war has begun.
deepseek is starting a price war on the ai market
Note that the new pricing is temporary as they are trying to increase their market share by giving a low price to try it. DeepSeek has said however that they expect lower prices in coming months as they build out their hardware.
One last note: Kimi 2.6 is still better at coding than V4 pro in tests I’ve seen and it’s not very expensive either - comparable to V4 pro. However Kimi is a highly distilled MoE of deepseek 3.2 base. So I expect that as people start distilling deepseek V4 we are going to see a lot of improvements in it – possibly in around 3 months.
In short, Chinese companies are killing US AI in the longer run. Unless Mythos level models come out and are cheap enough to use, the Chinese AI companies are going to get the dominant global share.
1 Like
Interesting short (but highly technical) talk on how Deepseek managed their new 1Million token context window:
1 Like