Mike Knoop
@mikeknoop
Followers
22K
Following
9K
Media
245
Statuses
4K
co-founder @ndea and @zapier @arcprize
sf bay area
Joined July 2009
> AI tools require adjustment in process. This strikes me as true. Effectively using AI in tools like Cursor or VS Code requires changing existing habits and coding patterns. You have to actively manage context and docs to get the best results.
I think one factor driving varied results is that AI tools are actually not that easy to use right away and require a learning curve of some hours (see @simonw). Another is that they require adjustments in process. A third is that AI has different uses depending on user expertise.
2
0
14
Zooming out ARC progress: I'd say OpenAI's o-series progression on v1 is a bigger deal than Grok's progression on v2 (so far). The o-series marked a critical frontier AI transition moment from scaling pretraining --> scaling test time adaptation. Whereas Grok 4 mostly takes.
17
23
323
This is accurate. We verified Grok 4 using our semi-private ARC datasets.
@BasedBeffJezos Important to note that ARC tested @Grok 4 independently to achieve those results. Those results are not from us.
395
1K
8K
Perhaps the most unintuitive thing about AI today is that AI can simultaneously score 50%+ on Humanity's Last Exam (relatively hard for humans) while only scoring 16% on ARC-AGI-2 (relatively easy for humans). Example v2 task below.
33
23
279
Grok 4 takes the #1 spot on the ARC v2 commercial leaderboard with an impressive jump over previous SOTA.
Grok 4 (Thinking) achieves new SOTA on ARC-AGI-2 with 15.9%. This nearly doubles the previous commercial SOTA and tops the current Kaggle competition SOTA
5
16
161
I think people in the US generally overestimate the traction a third national party needs to gain in congress in order to have significant negotiating leverage.
BREAKING: Republicans lose Rep. Brian Fitzpatrick, meaning if votes hold, the ‘Big Beautiful Bill’ will fail. GOP will attempt to flip someone.
1
0
36
RT @tony873004: Interstellar visitor confirmed. #A11pl3Z is now known as 3I/ATLAS. It is only the third confirmed object from beyond our s….
0
340
0
Assuming steady AI progress, the design goal for ARC v2 was to endure 12-18 months. Goal for v3 is >3 years. (Note this July 17 event is a preview launch of the first v3 interactive reasoning games to get feedback. We plan to ship full v3 dataset early 2026).
ARC-AGI 3 was just announced, the AI benchmark by Francois Chollet. This came unexpected because ARC-AGI 2 is very far from saturation at this point. @arcprize
1
11
56
The more idea constrained we are toward AGI progress, the more true this is. And evidence strongly suggests we are idea constrained. Making a big impact requires doing something very different than everyone else.
Its not too late to have impact on AI. It's not just up to the current leaders of AI. People who are motivated to have impact can get up to speed super fast. Worth repeating.
3
5
47
RT @fchollet: Impressive results from Sakana AI on ARC-AGI-2 with a new method for test-time-search and ensembling!. Please be mindful when….
0
140
0
If marketing continues to diverge from reality, the market demand for progress will go up. Good news for all of you "just" working on AGI.
Chat, I want to start a discussion:. Why did all the AI labs stop talking about AGI, and instead "superintelligence" is the new word?. It can't be coincidence.
2
1
12
Accuracy vs efficiency reporting is great to see.
We're taking a big step towards medical superintelligence. AI models have aced multiple choice medical exams – but real patients don’t come with ABC answer options. Now MAI-DxO can solve some of the world’s toughest open-ended cases with higher accuracy and lower costs.
2
0
9
Clarifying, this was the first time direct imaging was used to confirm discovery of a new exoplanet. First-ever direct exoplanet image from JWST was in 2022:
0
0
4
NASA JWST just reported its first-ever direct imaging of an exoplanet (infrared): TWA 7b. It's about 110 ly away, 50AU from its host star, and 30% mass of Jupiter.
1
1
23
RT @peterwildeford: New challenge for AGI: if AI could solve 60+ games never seen by AI or human developers. A really cool evaluation in de….
0
3
0
Tech-risk startups (eg. Ndea) and market-risk startups (eg. Zapier) share one thing in common: frequent contact with reality increases your chance of success by increasing your rate of learning.
Key to research success: ambition in vision, but pragmatism in execution. You must be guided by a long-term, ambitious goal that addresses a fundamental problem, rather than chasing incremental gains on established benchmarks. Yet, your progress should be grounded by tractable.
2
1
44
If you are a capable AI researcher with new ideas, please consider acting on them! Predicting breakthrough is hard because you don't know if it's the lack of pre-requisite ideas or merely action.
We asked @mikeknoop (co-founder of @ndea, @zapier, and @arcprize) about the limits of today’s AI models. “We are idea constrained to get to AGI. There are some major breakthroughs we have not figured out yet.”. “At the beginning of last year, there was a serious vibe of: 'It’s
8
4
31