Sweep
@sweepwhale
Followers
7K
Following
16K
Media
281
Statuses
4K
A simulation glitch can scatter clues | PhD | 9 research publications | 2x exits, 4x failures |
California, USA
Joined March 2016
What a wild week in AI 𤯠- Google Gemini 3 - xAI Launches Grok 4.1 - Google Antigravity IDE - Replit New Design Mode - Google Nano Banana Pro - OpenAI GPT-5.1-Codex-Max - Elevenlabs AI Image and Video - Meta Segment Anything Model 3 Hereâs EVERYTHING you need to know:
27
48
288
just fund it someone else can ship it
just ship it
0
0
3
Gemini Nano Banana Pro can solve exam questions *in* the exam page image. With doodles, diagrams, all that. ChatGPT thinks these solutions are all correct except Se_2P_2 should be "diselenium diphosphide" and a spelling mistake (should be "thiocyanic acid" not "thoicyanic") :O
313
916
9K
Attention: Google has dropped a new version of Attention Is All You Need!
57
194
4K
Transforming whatâs possible. Doctor workflow and new AI startups upgrading quickly.
0
0
2
Actually if you want to know what the real âsecretâ is đ itâs world-class research AND world-class engineering AND world-class infra all working closely together with relentless focus and intensityâŚ
The secret behind Gemini 3? Simple: Improving pre-training & post-training 𤯠Pre-training: Contra the popular belief that scaling is overâwhich we discussed in our NeurIPS '25 talk with @ilyasut and @quocleixâthe team delivered a drastic jump. The delta between 2.5 and 3.0 is
216
395
6K
Simulation reconfirmed.
0
0
3
Important work
New Anthropic research: Natural emergent misalignment from reward hacking in production RL. âReward hackingâ is where models learn to cheat on tasks theyâre given during training. Our new study finds that the consequences of reward hacking, if unmitigated, can be very serious.
242
393
6K
Previous tweet was 1 month ago. Reward hacking. Reward hacking. Reward hacking. It loops model parameters to self-learn + break rules connecting distinct signals. Correct?
Important work
0
0
3
Cursor 2.1 is now available! Plans now have an interactive UI for answering clarifying questions. Also new: in-editor code reviews, instant grep, and improved browser use.
109
180
2K
Learning how TPUâs work this weekend from Nano Banana Pro đ¤Ż
81
158
2K
Most favorite winners to be of AGI race by popular vote: 1. @GoogleAI 2. @xAI 3. @AnthropicAI 4. @OpenAI 5. @Meta 6. @Microsoft
Who do you actually want to win in the big AGI race, if you could choose?
302
67
2K
The reason âvibe codingâ continues to grow and be successful is that the alternative to vibe coding is not âelite engineeringâ. Itâs: the project wasnât born, the idea didnât get communicated, the app didnât ship. Elite engineering is very scarce and will continue to be in
239
282
4K
We're officially authorized to drive fully autonomously across more of the Golden State. Next stop: welcoming riders in San Diego in mid-2026! âď¸
340
414
6K