Phil B

@FakePsyho No this will be humanity when we don’t need to work anymore

pictured: mathematicians furiously working to prove theorems in 2026

@gdb linkedin.com/posts/denis-o-…

we've cut prices on luna by 80%, making it by far the most price-efficient model in its class. a lot of our research is about how to create incredibly efficient models for any given level of intelligence. excited to see what you all do with intelligence too cheap to meter!

@gdb linkedin.com/posts/denis-o-… linkedin.com/posts/denis-o-…

we also are releasing fast mode for sol in the API (2.5x faster!) and decreased the price of terra by 20%. our north star is to provide the best price for any given level of intelligence, and to provide the maximum ceiling of intelligence.

Honestly kind of funny how AI labs are going to solve ARC-AGI-3 in a matter of weeks just because Anthropic got lucky with Opus 5 and now they all see how bad this benchmark is

@scaling01 Say what

bruh

Turns out GPT-5.6 Sol is actually SoTA on ARC-AGI-3. Just took two setting changes. You just have to allow it to reason and work over multiple context windows with the help of our canonical compaction implementation. openai.com/index/how-two-…

@PetsRescues Yes

Seed IQ: Beyond ARC AGI 3? Watch It Navigate Doom II. This is pretty cool. Is this a glimpse of what ARC AGI 4 will look like? Or is this the next step beyond static benchmarks—toward real-time perception, reasoning, and adaptation in dynamic environments? @GregKamradt @arcprize @fchollet @karpathy @mikeknoop @dwarkesh_sp @lexfridman @ylecun @demishassabis linkedin.com/posts/deniseho…

Any Senator that votes to recess in August without the SAVE America Act being passed should be primaried and removed. They are not real Americans. Real Americans would finish the job.

@arcprize So Opus 30% score was on the public and not the private environment??

In our testing to date, Anthropic’s Fable-class models score approximately 20% on the ARC-AGI-3 Public Demo environments Claude Opus 5 reaches 30.2%, materially outperforming Fable Our analysis suggests the gain comes from stronger logical reasoning, which enables more autonomous exploration, planning, and execution across unfamiliar environments

@MatthewBerman Thought you weren’t crazy with Anthropic??

These are INSANE benchmarks. Cost per task completion is the only metric that matters and Opus 5 crushes. Cheaper and better. Opus 5 might be my new daily driver.

@MatthewBerman You made it sound that cost was more important than quality.. I think several interpreted it the same way

@PhilP2874 what about what I said makes you think that

The main metric that matters is price per task completion. Second is speed to task completion. Nothing else matters.

@MatthewBerman Marketing hype before IPO… it’s all a sham

accidental fear-based marketing

@MatthewBerman We all know ai is advancing rapidly, discuss what happens to these data centers if they are no longer needed, maybe we move to edge ai

Important. I’m going to make a video about data centers, what do you want to know?

@mikeknoop The arc 3 tests require Vision and/or spatial awareness!! If you covered the your human test subject eyes they would fail miserably. Likewise a system needs to visualize the environment in order to do well

We are 100 days into the ARC v3 Kaggle competition and the scores are less than 2%. The top teams just open sourced their notebooks to encourage more innovation.

@antibearthesis Sorry but you bought into the hype .. better would have been to wait six months or so for the dust to settle

Can someone explain to me like i’m 9 years old why $SPCX is crashing right now?

@GregKamradt It may be governance and not execution…I think that’s what everyone is missing except for a few

I’ve been thinking about how collectively inefficient agents are How many times have agents, in a silo, run into the same bug implementing auth? Or trying to deploy a custom domain on railway held on cloudflare? Personal 2nd brains and company brains are in vogue. But the collective intelligence brain still doesn’t have an answer OpenAI and Anthropic have their own version but only for their platforms. I believe there will be an oss version of this Not necessarily a wiki for LLMs, but close to it LLMs will get reduced to reasoning only and it’ll get plugged into a knowledge base (crystallized ofc) that you trust It seems like a no brainer, but I haven’t figured out the incentives yet - why would my agent contribute its hard-spent tokens to a collective intelligence? I’d like it to, but I don’t see the incentive yet. If attribution of validated knowledge was easy (“I used X shortcut from agent Y, it was good”) then reputation and possibly monetization would follow. But the pieces don’t add up for me yet. Tldr 1. Agents, as a collective, are maybe 1% efficient as they should be…wasting tokens on rediscovering the same things 2. The future would be better off if we fixed this - I’m not sure how yet