Einsia ✓

@EinsiaAI

We're teaching AI the work the world's experts actually do.

585Followers

6Following

70Posts total

2.1MViews on collected posts

Against accounts of the same size

7 posts from the last 90 days, next to the under 10K follower range. reaches fewer people than peers, but engages them much harder.

Median views565this account4 164median for under 10K

Reach, %96.58%this account277.56%median for under 10K

Engagement, %2.48%this account1.20%median for under 10K

Metric	This account	Median for under 10K	Ratio
Median views per post	565	4 164	0.14×
Reach (views ÷ followers)	96.58%	2.8× audience	0.35×
Engagement rate	2.48%	1.20%	2.07×

Others in this range → Compare with another account → How these benchmarks are built →

Growth & engagement

How the posts we collected actually performed: views and reaction rate post by post, what the audience did with them, and where the follower count goes.

Views per post

2.1M21 Aug

1.8K

799

565

498

442

128

Last 7 collected posts, oldest on the left. The scale is logarithmic: one post can outrun the rest a hundred times over.

Engagement rate per post

0.05%21 Aug

1.06%

2.25%

2.48%

3.01%

3.85%

7.81%

Reactions — likes, reposts, replies and quotes — divided by views. Median for under 10K accounts is 1.20%.

What the audience does

Likes48.5%753 in total

Reposts8.6%134 in total

Replies7.9%123 in total

Quotes8.4%131 in total

Bookmarks26.6%413 in total

Share of every reaction we collected for this account. Replies mean argument, reposts mean endorsement, bookmarks mean the post was worth keeping.

The follower curve appears once this account has two daily snapshots — we take one a day, and this one is on its first.

Latest posts

@EinsiaAI We should not be working on RSI or research that could enable RSI. By default RSI probably results in AIs that are uncontrollably smart and are misaligned with human interests. Such AIs likely kill all humans. 128 views · 10 likes · 0 reposts · 0 replies 21 Aug 2026 6/ The takeaway: Today’s agents can only occasionally do real algorithmic design—not reliably. Even at the highest reasoning tier, the average score was only 0.196: about 11% of the way from the 0.10 baseline to the theoretical optimum of 1.0. The question isn’t whether agents 442 views · 16 likes · 0 reposts · 1 replies 21 Aug 2026 5/ More reasoning means more exploration: 13× more code, 10× more tokens, 4× more evals. But gains lag behind: score improves only ~2×, reaching just 0.196 at max effort. https://t.co/SfajILjlWp 498 views · 13 likes · 0 reposts · 2 replies 21 Aug 2026 4/ What agents changed mattered most: Of 263 submissions that made a change: -141 changed only the run settings -122 changed the learning method itself In terms of score, 0.226 for learning-method changes vs. 0.126 for run-settings changes. Changing algorithms performed ht 565 views · 13 likes · 0 reposts · 1 replies 21 Aug 2026 3/ Across 29 model × harness × effort configurations: Scores range from 0.065 to 0.288, with an average of just 0.166. Opus 5 is the best. GPT, Claude, and Kimi stay fairly close overall. Median exploration cost per task jumps from $1.69 to $34.60—but spending more doesn’t htt 799 views · 16 likes · 0 reposts · 2 replies 21 Aug 2026 2/ AI4AI-Bench: 10 research repos, spanning 10 algorithm families. Agents get the repo, starting artifacts, and a proxy metric, then 4 hours on one B300 to improve the training algorithm. They submit code only—no trained weights or cached state. The code is rerun from scratch 1.8K views · 18 likes · 0 reposts · 1 replies 21 Aug 2026 1/ Recursive self-improvement (RSI) depends on agents improving how AI systems are trained —not just tuning hyperparameters, but improving the training algorithm itself. We tested this directly with AI4AI-Bench: 10 real research repositories spanning 10 distinct algorithm https 2.1M views · 667 likes · 134 reposts · 116 replies 21 Aug 2026