Chinese AI is already superhuman when it comes to silly walks. https://t.co/MkQJKS0C8k
— Will Knight (@willknight) 20 avril 2026
Chinese AI is already superhuman when it comes to silly walks.
RESEARCH
-
Chinese AI Achieves Superhuman Performance in Silly Walks
By
–
-
Snorkel Workshops on Data Quality and Agents at ICLR
By
–
Big week ahead at @iclr_conf Brazil🇧🇷
— Snorkel AI (@SnorkelAI) 20 avril 2026
Don’t miss where Snorkel will be from workshops with @fredsala and @vincentsunnchen on data quality and agents to a @DFAI_Community Gathering on Saturday.
Full lineup below👇 https://t.co/ISe7mjN7kOBig week ahead at @iclr_conf Brazil Don’t miss where Snorkel will be from workshops with @fredsala and @vincentsunnchen on data quality and agents to a @DFAI_Community Gathering on Saturday. Full lineup below
-
AI to Add 0.6% Annual GDP Growth Over Next Decade
By
–
They all predict around 0.6% additional growth PER YEAR over the next decade due to AI.
In other words, if GDP growth without AI was going to be 2.5% per year on average, it will now be about 3.1% per year. The compound growth over 10 years will be 36% instead of 28%.
The GDP -
Multimodal LLMs Fail at Car Detection Task Compared to YOLO
By
–
I tested Opus 4.7 on a simple car detection task.
— Satya Mallick (@LearnOpenCV) 20 avril 2026
→ Took 5 minutes
→ Missed multiple cars
→ Pointed at blank spaces
→ Bounding boxes were even worse
Codex did better (24 cars, accurate points) but still took 3 minutes.
YOLO does the same thing in 30ms.
Multimodal LLMs are… pic.twitter.com/DiHGsYpOG2I tested Opus 4.7 on a simple car detection task.
→ Took 5 minutes
→ Missed multiple cars
→ Pointed at blank spaces
→ Bounding boxes were even worse
Codex did better (24 cars, accurate points) but still took 3 minutes.
YOLO does the same thing in 30ms.
Multimodal LLMs are -
Low Rank Dense Bubbles in Machine Learning
By
–
yea i think it's more about "low rank" here being close to the centroid of a local dense bubble
-
1D Ordered Tokens Improve AR Image Model Test-Time Search
By
–
“(1D) Ordered Tokens Enable Efficient Test-Time Search” Most AR image models generate 2D grid tokens patch-by-patch, so partial generations are hard to verify. This paper shows that 1D coarse-to-fine tokens make search much easier, with early tokens encode global semantics,
-
AI Race Nowhere Near Final Sprint Yet
By
–
Sorry, Sergey, but we're nowhere near the final sprint in the AI race.
-
Latest AI News from OpenAI, Anthropic, Microsoft and Perplexity
By
–
P.S. — I won't be at #GoogleCloudNext, but @Kantrowitz will be! Today's newsletter also includes a look at recent AI news from @OpenAI
, @AnthropicAI
, @Microsoft
, @perplexity_ai
, etc. It also has some of the stats in @StanfordHAI
's 2026 AI Index. -
Memory as AI’s Critical Bottleneck: Open Access Imperative
By
–
memory will be the great lock in
— Harrison Chase (@hwchase17) 20 avril 2026
and everyone knows it, so are rushing to get there first
memory should be open! https://t.co/uKDq4ChMWNmemory will be the great lock in and everyone knows it, so are rushing to get there first memory should be open!
-
Free AI Training Access Despite Data Center Concerns
By
–
People can feel how you want about data centers while still taking advantage of free training that opens up more options for them.