Excited about AI and sharing thoughts
Beit El
Joined August 2009
- Tweets2.9K
- Following399
- Followers178
- Likes5.4K
True.
I feel lucky to be Israeli.
Out of 14m Palestinians in the Middle East, Arab Israelis are the *only* ones that get a vote.
Palestinians in West Bank haven't voted since 2006, stuck under a 90 year old dictator.
Lebanese Palestinians don't even get the citizenship.
Jordanian Palestinians have a.........king.
You may not like this, but the best thing to happen to Palestinians is to live inside Israel.
Eat your heart out.
46
10/ Whack-a-Mole, 3 one-minute rounds: Easy (3x3, ~1.2s window): 31/33 (94%) Medium (4x4, ~0.9s): 54/60 (90%) Hard (4x4, ~0.6s, decoys): 51/59 (86%) Random clicks: 9, 6, 5 ~170ms per call, ~0.3-0.4s from pop-up to whack. ~620 frames ≈ 4¢.
1
26
15/ Honest limits: • Dense crowds with look-alike decoys: sol is far ahead (83% vs 52%) • Tiny targets ("+", page numbers) get missed • `box` always returns a box, so pair it with a yes/no when the target may be absent • Image inputs are still behind a danger-zone header
1
1
20
16/ Bottom line: Jev 1.14 `box` is a ~0.2s, sub-cent grounding call with confidence you can actually threshold on. Enough for real-time agents: click bots, game loops, visual QA. Not the most accurate model out there, but the fastest one I can trust. #Jev @typesafeai
31
8/ The part I like most: Jev knows when it doesn't know. Across the Waldo scenes, when confidence was ≥0.7 it was right 79/79. And it stays fast: ~0.2s median even with 600 people in the frame (~$0.0002 per scene).
1
35
3/ Real-looking UIs (admin panels, shops, editors, settings): 24/28 (86%). Every Hebrew instruction hit. The misses were tiny targets like a "+" button or page "2". ~0.2s per question, ~$0.0001 each.
1
21
6/ Where's Waldo 🔴⚪ Jev gets ONE reference image of Waldo + a scene of 30-120 people. But Waldo never matches the reference: lying, upside down, sitting, back turned, shrunk, behind trees/people, only his head peeking. Result: 70/72 (97%) in one call.
1
20
5/ Then I removed text entirely: 54 symbol/drawing tests, 50 hit (93%). App grids 8/8, emoji faces 5/5, drawn maps 8/8, "smallest star"-type 4/4, toolbars 14/16, shape scenes 7/8, odd-one-out 4/5. Avg 0.18s. Whole batch < $0.003.
1
9
4/ Favorite one: "undo the last change" on an icon-only toolbar. Jev's confidence was just 0.52, and it was still right. 0.21s, ~$0.00009. Low confidence ≠ wrong, but (spoiler) high confidence almost always = right.
1
8
Jev found Waldo in 0.2 seconds! I stress-tested the new `box` question in @typesafeai #Jev 1.14: clicking UIs, finding Waldo in crowds of 600, playing Whack-a-Mole in real time, and racing OpenAI. All numbers below are from my own runs 🧵
1
93
Aviz Maeir retweeted
Meet Vidu Q4 Preview—our next-gen flagship AI video model.
Expressive performance. Cinematic camera language. High-impact VFX.
Vidu Q4 Preview is now live. 🔥
Flagship AI video, made accessible— from $0.014/sec.
Try now with new user invite code VIDUQ4PV1
Follow + RT + Reply this post = 200 Credits in DM (72 Hours Only)
🤖 Made with AI
