ceo @hillclimbai / ex pro valorant @misfitsgg evals @googledeepmind
de_mirage
- Tweets 17,544
- Following 1,825
- Followers 11,527
- Likes 32,002
Introducing DeepRSI, a new industry standard for measuring recursive self-improvement.
On preexisting benchmarks, tasks are contributor-based and short-horizon (~12 hrs). In DeepRSI, we push agents to work 48 hours on ML tasks actually derived from ongoing proprietary research.
184
192
1,704
223,607
some days you feel off but don’t know why
introducing the world’s first non-wearable fitness tracker, built for your mind
Soma reads 200+ signals from your voice to show you when you’re sharpest and what slows you down
it runs alongside meetings, voice notes, and dictations to bring you insights when they matter most
make your best days your everyday, join our private beta now: soma-health.co
67
56
480
246,637
Sales tax is coming. And we are giving away California registrations for free.
Starting January 1, 2027, California will tax most software. For many founders and teams, this means collecting, filing, and remitting for the first time.
For a software business, sales tax has typically been something to worry about at, let's say, the Series A, or when your business is getting close to eight figures in ARR.
But SB 122 changes this logic. The risk is really under-collection, or lack of collection, creating a future liability for the company.
@Numeral is doing two things through the end of the year:
- we will do a free tax analysis
- we will handle your California registration for free
The offer is open to any software companies that aren't yet Numeral customers.
DM @NateMatherson. He will get you squared away.
19
6
41
3,779
hello just in case its not obvious:
i’m nowhere near as knowledgeable as john schulman, karpathy, or many other researchers on RSI
it just so happens that i am incentivized to make RSI as legible of a concept as it can be to others and i do that well
that being said if you wanna chat bout RSI, my DMs are always open 😄
I asked 51 people in AI who has a good read on RSI (recursive self-improvement) and where it's going:
5 votes:
- Jerry Tworek, Core Automation (@MillionInt)
4 votes:
- Jun Park, Hillclimb (@jparkjmc)
3 votes each:
- John Schulman, Thinking Machines (@johnschulman2)
- Rohan Anil, Core Automation (@_arohan_)
- Daniel Chang, ex-Mercor (@dnnssl2)
- Laksh Aithani, Abundance (@LakshAithani)
- Richard Socher, Recursive (@RichardSocher)
2 votes each:
- Andrej Karpathy, Anthropic (@karpathy)
- Behnam Neyshabur, Mirendil (@bneyshabur)
- Han Fang, Meta Superintelligence Labs (@Han_Fang_)
- Herbie Bradley (@herbiebradley)
- Jennifer Zhou, Preference Model (@chem_safety)
- Paul Christiano, Alignment Research Center (@paulfchristiano)
75 people were picked. The other 62 got 1 vote each and are in the next posts.
51 people voted, each picking from people they've met. I didn't vote.
10
1
134
8,800
what happens to openai there?
Gentle Monster is taking 28,500 SF at the Puck Building in SoHo.
- 295 Lafayette St, owned by Kushner Companies
- Mostly the former REI flagship space
- Lease runs more than 10 years
- Kushner says the deal completes the full re-tenanting of the building
The South Korean eyewear brand is known for gallery-style flagships. Its current SoHo store is at 70 Wooster St.
2
39
11,945
my dawg got ycombinator posting about bed wars @sensho @olam_labs 🤣🤣
Next, we ran nine 45-minute games of Bed Wars, a Minecraft game format where teams fight to destroy each other’s beds while protecting their own.
We pitted 10 Claude Opus 5.5 agents against 10 GPT 6.1 Sol agents with the instruction: "win the game." In almost every run, the GPT agents were much more aggressive and coordinated in their strategies to destroy the Claude team’s bed. Team GPT won the series 6-3.
3
1
76
4,635
"roll out"
Have been working on task extraction for the past month, and the diagram below is actually fairly accurate!
That said, this is like a modern day fusion reactor — will the money coming out be good enough to sustain the money that needs to go in?!
It's surprisingly costly to generate high-quality tasks. For each task you run what's called a "roll out", essentially a single turn by a given model + reasoning level. It might take you 100s of rollouts across various models until you can have a good sense of the task, and each of those runs cost $$$!
34
1,643