Everything is fishy about this company
We just served 1 trillion tokens of Beam in 24 hours. To celebrate, we've reset everyone's limits.
Go build something big 👉 platform.reflection.ai
Somehow moving toward a feeling that Sol 6.1 is better at orchestrating long-running work.
What's very paradoxical is that opus is much more steerable but somehow loops into mistakes
So what if we include another model weights in the pretrain of an LLM, any interesting properties could emerge?
what happened to deep research? Everyone forgot about it. Seems like we fully moved to agentic dynamic workflow outputs
Agi benchmark:
When planning architecture and data model (or other verifiable task) can a model with /goal autonomiously outperform you againts you iteratively working with the model? If yes it means you are the bottleneck for exceptional results not the model, and AGI is here.
So far human in the loop work vastly outperforms autonomious one where exceptional coherence is required. We have not reached a point where the entropy is getting minimised with long horizon work
Here is the reason why google market cap is not 8T right now, why models forever behind frontier and cross product experience so inconsistent.
Crazy to think about how much latent potential is lost when company is not aligned with itself
I was on the Docs team 15 years ago. The rewrite (from Writely to the new OT-based editor) had just launched. We were facing an enormous list of features we lacked that prevented enterprises (and some consumers) from using us. I think we didn't even have inherited text styles at that point. Markdown was so far down the list compared to absolute basic editor functionality. Plus there were some reasonable arguments against it, like "the 'spec' is some guy's webpage"
The thing that stopped it for the next 5 years (which was how long I was on the team) was that the Word <-> Docs conversion service was a separate team based in Sydney called Changeling. Having literally opposite work hours made it difficult to coordinate with them, plus it made it easy for the I-don't-want-Markdown-in-Google-Docs crowd to just say "obviously Markdown conversion should be handled by this other team that's really difficult to coordinate with," which was a really effective way to shut it down in favor of things that would better help them game promotion.
I don't have any insight into why they finally started prioritizing it a few years ago, but my guess is that they started losing contracts to Notion which is a lot more Markdown-native
Borouch retweeted
Using arXiv publication trends, we could estimate when AI might provide as much uplift to a given field as it does to mathematics today. Quantum physics could reach that point within a year, and physics as a whole by late 2029. Biology, of course, hasn’t budged yet.
Having no shame or cringe factor in social space is like a 100x leverage multiplier for your life. Network effects, marketing, new relationships, opportunities. Participating in the attention economy, in most cases, far outweighs the negatives.
Incredible parallels to the recent ai alignment incidents
Incredible emergant capabilities.
Ofc labs figured out that to win memetic war of model release it needs to be beautiful, impressive and new. So they will keep RL training new visual forms like this to capture attention.
The biggest bangers to be expected before IPOs
Human brain learning efficiency if had the capacity, playing off againts gpt-4 era pretrain data volume should already easily achieve super intelligence.
Current models are order of magnitude below human learning efficiency
We are just brute forcing our way through to agi rn
