Your curated collection of saved posts and media
Had early access to GPT-5.2. Its an impressive model. Here is GPT 5.2 Pro's version of "create a visually interesting shader that can run in twigl-dot-app make it like an infinite city of neo-gothic towers partially drowned in a stormy ocean with large waves," single shot. https://t.co/ZLeXZ7OIIn
In the more practical example: "build me a graph of humanity's last exam scores over time" which involved looking up and cross-referencing a lot of material and then generating something useful in one shot: (Ironically does not include GPT-5.2 since scores weren't public) https://t.co/FmeyzzdGHP
After reading it, this does seem like a big deal Industry experts outlined important, real-world, hard tasks for AI to do. Other experts were asked to do the tasks themselves & yet others graded human & AI output Models approached parity with humans & AI is getting better fast. https://t.co/z666YcNyH6
Whoa. This new GDPval score is a very big deal. Probably the most economically relevant measure of AI ability suggesting that in head-to-head competition with human experts on tasks that require 4-8 hours for a human to do, GPT-5.2 wins 71% of the time as judged by other humans https://t.co/M8NqSUXl6X
After reading it, this does seem like a big deal Industry experts outlined important, real-world, hard tasks for AI to do. Other experts were asked to do the tasks themselves & yet others graded human & AI output Models approached parity with humans & AI is getting
Note that 5.2 Thinking gets a lot of ties to get above the 50 % mark - but 5.2 Pro has a 10 % point lead on pure wins vs 5.2 Thinking, even if the total win&tie-rate ends up being "only" 3,2 % higher. Clearly 5.2 Pro delivers more robust economically valuable quality. https://t.co/mRIloK299p
๐ A huge thank you from #AWS to the #startups, speakers, experts & innovators for creating an unforgettable #awsreinvent. โจ We shared insights, discovered new possibilities & left feeling inspired by the brilliant minds shaping the future of technology. https://t.co/aEY6rrjRug https://t.co/72uMrvdFkS
12ๆ23ๆฅ็ซๆๆฅใซSakana AI ้่รAIๅๅผทไผใ้ๅฌใใพใ๏ผ๐๐๐ ้่ๆฅญ็ๅ จ่ฌใฎๅค้ฉใซ้ขๅฟใฎใใใใจใณใธใใขใจใใธใใน่ท็จฎใฎๆนใๅฏพ่ฑกใจใใฆใใพใใ โข ้่้ ๅใซใใใๆๆฐใฎ็ฅ่ฆใจใใใฎ็คพไผๅฎ่ฃ ใใญใปใน โข ใฐใญใผใใซใช็ฐๅขใงใไธ็็ใช็ ็ฉถ่ ใปใจใณใธใใขใจๅ ฑใซๅใ้ญ ๅ โข ๅฎๅ จๆงใๆฑใ ใใใ้่ใทในใใ ใธใฎAIๅฐๅ ฅใซใใใ้็บไฝๅถ ใชใฉใใใใงใใ่ใใชใๅ ทไฝ็ใชใใจใไผใใงใใใฐใจๆใใพใ๏ผ 19ๆ30ๅใใใใใฏๆ่ฆชไผใๅฎๆฝไบๅฎใงใใๅฝๆฅใฏใSakana AIใงๅใ้่ใใญใธใงใฏใใๆ ๅฝใใใจใณใธใใขใใใญใธใงใฏใใใใผใธใฃใผใชใฉใๅคๆฐๅๅ ใใพใใ ้่ๆฅญ็ใฎๆชๆฅใAIใงๅ ฑใซๅตใใใจใซ่ๅณใใๆใกใฎๆนใฏใใใฒใๆฐ่ปฝใซใ่ถใใใ ใใใ ่ฉณ็ดฐใฏใใกใโ https://t.co/HsdEXwdC7Z
It is a very smart model, and we have come a long way since GPT-5.1: https://t.co/6FJG5FbOQG
We at @MeritechCapital are excited to be in business with @getserval. For too long, IT teams have been held back by static, brittle ITSM solutions that prevent themโand their employeesโfrom performing at their best. AI-driven automation was always the promise, but no one delivered....until Serval. The company has grown revenue by 500% since closing its Series A only a few months ago and is used by some of the most forward-thinking IT teams today, including Clay, Together, Verkada, and many others. The company is also led by an exceptional team and CEO, @jakeserval. We're excited to be a small part of the future of ITSM with the Serval team. https://t.co/A62Bt8yuLA
Livestream begins at 12pm ET. Tune in, we'll be sharing in Real-time. https://t.co/eddEUQnvn9
Part 2 of @runwayml 4.5 out now! With prompts this timem. Forgot to enter them in the main post last time. Prompts used: - hand held shot of a baby lion waking up next to his mom's face. Focus on the textures. - First-person view of parkour runner's perspective. Sprint toward cliff edge, explosive takeoff, hands reaching out toward hot air balloon basket approaching fast. Balloon fills frame as jumper flies through air. Ground drops away far below. Adrenaline rush captured in motion. GoPro-style dynamic movement. - macro shot of sloths claws picking up a blade of grass -drone shot of bison crossing the nile river. Smooth cinematic motion #research #creative #researchanddevelopment
Deep Research is the first agent released on the new Interactions API โ offering a single endpoint for agentic workflows. Start building today โ https://t.co/FlgtzbDYj7
NEWS: Rivian haศ unveiled its Autonomy chip and Gen 3 Autonomy Computer, which the company says is designed to solve the needs of autonomous driving. Chip: โข Mutli-chip module โขย TSMC 5nm โขย Neural engine: Rivian designed โขย 800+ TOPS โขย In-house designed software stack "RAP1 powers the companyโs third-generation Autonomy computer, the Autonomy Compute Module 3. Key specs include: โขย 1600 sparse INT8 TOPS โขย The processing power of 5 billion pixels per second. โขย RAP1 features RivLink, a low latency interconnect technology allowing chips to be connected to multiply processing power, making it inherently extensible. โขย RAP1 is enabled by an in-house developed AI compiler and platform software.โ Rivian: โIn addition to ACM3, Rivian plans to integrate LiDAR into future R2 models. LiDAR will augment the companyโs multi-modal sensor strategy,ย providing detailed, three-dimensional spatial data and redundant sensing, and improving real-time detection for the edge cases of driving. Our Gen 3 Autonomy hardware including ACM3 and LiDARย is currently undergoing validation and we expect it to ship on R2 models starting at the end of 2026.โ
OpenAI says GPT-5.2 Thinking hallucinates less than GPT-5.1 and has improved reliability for agentic AI needs; pre-release testers include Notion, Box, Shopify (@haydenfield / The Verge) https://t.co/I64lWbO1Om https://t.co/iLX7mjbATR ๐ฅ Send tips! https://t.co/wlNZvXuhJs
This seems Huge Scientists at Caltech and Cedars-Sinai have built a new AI tool called NOBLE that can quickly and accurately create virtual versions of brain neurons. Basically, it helps researchers understand how the brain works a lot faster, which could eventually lead to better treatments for brain related disorders.
A team of scientists led by Caltech and Cedars-Sinai has developed a new artificial intelligence framework that can accurately, quickly, and efficiently create virtual models of brain neurons. https://t.co/mxRpp2jvTM
GPT-5.2 Thinking evals https://t.co/Kcnz3ZIwye
GPT-5.2 Thinking evals https://t.co/Kcnz3ZIwye
WonderZoom Multi-Scale 3D World Generation https://t.co/6VAoAzZ4SO
discuss: https://t.co/FKh81wKg5f
Apple presents Learning Unmasking Policies for Diffusion Language Models https://t.co/UGSQIdxxB5
discuss: https://t.co/kCLhzs6dPO
๐คฏMUST TRY: Qwen-Image-i2L skips the training loop entirely. 1-5 images in โ LoRA weights out in seconds. โฌ๏ธ Demo available on Hugging Face https://t.co/yuqEZ1GAnw
Towards a Science of Scaling Agent Systems https://t.co/JmCe3VLTby
discuss: https://t.co/uJHaaQkg5G
We just released TRL v0.26.0! It comes packed with updates: > Agent training with tools in GRPO > New CISPO & SAPO losses > Reasoning rewards > vLLM quantization in colocate mode > Dataset shuffling in SFT > Lots of NEW examples > Tons of fixes and documentation improvements https://t.co/Vt3dmI1sLU
Sharing the slides from a talk I gave this week on bridging the gap between research experiments and building production-ready models, based on our recent Smol Training Playbook. https://t.co/RmG53PytMv
We are now the #1 trending text-gen <256B size model on HuggingFace!! https://t.co/SyoOHjWfvH
We are now the #1 trending text-gen <256B size model on HuggingFace!! https://t.co/SyoOHjWfvH
๐ฅ Ultra-FineWeb-en-v1.4 is coming! 2.2T tokens fully open-sourced! The core training fuel for MiniCPM4 / 4.1, fully updated based on FineWeb v1.4.0: ๐ What's New 1๏ธโฃ Fresher Data: Added CommonCrawl snapshots from Apr 2024 - Jun 2025 to capture the latest world knowledge. 2๏ธโฃ Easier Access: CC Dump Slices are here! No need to download the entire massive dataset anymore, fetch exactly what you need seamlessly. โก Highlights & Performance - Efficient Verification: Efficient Verification Strategy: Reduces data verification cost by 90% - High-Efficiency Filtering Pipeline: Optimizes selection of both positive and negative samples - Performance Gains: +3.613/+1.331 (Eng) & +1.98/+0.61 (Chn) vs. FineWeb/FineWeb-edu & Chinese FineWeb-edu-v2. Still high-quality cleaning. Still true to the open-source spirit. Welcome to download and test! ๐ ๐ Resources ๐ค Dataset: https://t.co/KluL5t2kUn ๐ Paper: https://t.co/Kg9LLUqZgB ๐งฉ Classifier:https://t.co/oUfxrN6AmP ๐ค MiniCPM4:https://t.co/IQ82jD1PTi #UltraFineWeb #MiniCPM4 #AI #LLM #OpenBMB #UltraData
> llama-cli -hf org/model
> llama-cli -hf org/model
llama.cpp gets a new CLI (tested it and it's ๐ฅ) https://t.co/XKVicocKGC
