Generated For AI @GeneratedForAI

AI News, Tools & Tech breakthroughs Joined January 2020

Tweets

908
Followers

61
Following

86
Likes

860

Generated For AI @GeneratedForAI

18 hours ago

arxiv.org/pdf/2610.20518

0 0 1 9 0

View Details

Generated For AI @GeneratedForAI

18 hours ago

A team of specialist AI agents can outperform one outstanding AI agent. Princeton and UC Berkeley tested this on GAIA, a benchmark for complex assistant tasks. The specialists scored at most 35% alone. Together, they reached 70%. The best individual system tested reached 67%. The idea is simple: let each developer build an agent around a specific skill, then connect their strengths. Web research, document analysis, working with images and audio. Here's how it works. The researchers used AI to simulate developers building 54 specialist agents. A coordinator breaks a request into smaller tasks. A router sends each task to a suitable specialist, learning from past results who's good at what. The coordinator combines their replies and requests more work when needed. And here's the craziest part. In a separate experiment, an agent scoring just 19.1% alone helped lift an existing system from 40.4% to 53.2%. Its overall score looked weak. Its specific skill filled a gap.

1 1 3 47 1

View Details

Generated For AI @GeneratedForAI

18 hours ago

Elon Musk says X just changed how your timeline decides what to show you. Here’s how the change works. Two updates went in: 1. A new AI ranking model. It predicts how much time you’ll probably spend on each post, and does it “much better” than before. 2. A retrieval model trained to surface posts people read all the way through. In plain English: one decides the order of your feed, the other decides which posts get into the pool at all. And here’s the catch. These numbers are X’s own. There’s no published method and no independent check. “User active seconds” measures time spent, not whether people were glad they spent it. Still, the chart shows 2026 running above 2023, 2024 and 2025 all year. The feed is being tuned for posts people finish. Whether that’s good for readers is a different question.

Elon Musk @elonmusk

24 hours ago

Two important changes to timeline recommendations: - New AI/SI ranking model that predicts probable time spent much better - Retrieval model trained to surface posts people read all the way through User active seconds is growing exponentially and hitting all-time highs almost

2K 2K 18K 2.7M 2K

0 1 1 31 1

View Details

Generated For AI @GeneratedForAI

18 hours ago

@dom_lucre punishment from god or just an unfortunate coincidence?

0 0 0 835 0

View Details

Generated For AI @GeneratedForAI

18 hours ago

@cb_doge makes sense, if people feel like thhe future is getting worse, having kids feel like much harder decision to take

0 0 0 31 0

View Details

Generated For AI @GeneratedForAI

19 hours ago

@elshayib_ that would be crazy if it actually happenes, but porting a game that big will probably be way harder than just throwing the AI at it

1 0 2 31 0

View Details

Generated For AI @GeneratedForAI

19 hours ago

@XFreeze yeah once people stop feeling the safety in their own homes and space, they always procastinate or think badly of theselves. It also impacts their thinking pattern and brain

0 0 1 76 0

View Details

Generated For AI @GeneratedForAI

19 hours ago

@cb_doge great, now I might get even lesser audience

0 0 0 21 0

View Details

Generated For AI @GeneratedForAI

3 days ago

Watching this clip from 2018 feels surreal. Elon casually explaining rocket thrusters on a Tesla Roadster to Joe Rogan. What’s something he’s saying today that you think could actually happen within five years?

0 1 1 54 1

View Details

Generated For AI @GeneratedForAI

3 days ago

@XEthanai @grok @karatademada @brivael @dvorahfr this is so saddening. people are hurting other people

0 0 0 59 0

View Details

Generated For AI @GeneratedForAI

3 days ago

@XEthanai cant believe you actualy did, thanks alot 🩶❤️

0 0 0 10 0

View Details

Generated For AI @GeneratedForAI

3 days ago

@MrBeast yeah you are definitely gonna get it

0 0 0 7 0

View Details

Generated For AI @GeneratedForAI

3 days ago

@visegrad24 this is so heartbreaking, I hope all this monkeys get punished by law or deported

0 0 0 884 0

View Details

Generated For AI @GeneratedForAI

3 days ago

Solar-panel inspection AI just got 5x smaller and more accurate. The new model has just 440,000 parameters and spots cracked glass and hot spots on solar panels from drone thermal images. Solar panels left outdoors pick up defects: cracked glass, things blocking the surface, and hot spots. These cut power output and can even become a fire risk. Inspections are moving from people walking the rows to drones with infrared cameras. The AI has to keep up, ideally on small, cheap hardware. Here's how it works. It starts from YOLOv8, a popular real-time object detector. The feature-extraction layers are swapped for "Ghost" layers. These compute a few features properly and generate the rest with cheap shortcuts. The middle of the network gets lighter layers too. An attention module weighs features at different zoom levels, because defects range from tiny hot spots to large covered areas. Then the model is pruned. During training, it's pushed to make unimportant channels useless, and then 80% of its channels are cut. And here's the craziest part. Pruning broke it. Its main accuracy score fell from 0.763 to 0.643. So the researchers had the original, unpruned model teach the slimmed one. It recovered to 0.768, slightly above its own teacher on that score, at a fifth of the size. The results: - 0.768 vs 0.672 for the YOLOv8 model it was built on, and the top score among the 13 detectors tested - 81% fewer parameters and 87% less computation than that YOLOv8 model - Faster R-CNN, with 40 million parameters, scored 0.566 - It found 75% of all defects in the test images The catch: it was never actually run on a drone or small device. The authors say speed, latency and memory weren't measured. The test set was also just 131 images.