Your curated collection of saved posts and media
We have a new best open source model to close out 2025: GLM 4.7! It's been ~6mos since the first closed source model, Opus 4, broke 73% on SWE-Bench and GLM does 73.8%! It's fantastic at math and coding and beats DeepSeek / Kimi. Very cheap at $0.6/M in, $2.2/M out, 200k context, and fast too: 70tok/s! The bigger story here is how the gap between closed and open source models has evolved. This time last year, OpenAI launched o1-preview and the closest OSS thing we had was DeepSeek R1-lite. We're still ~6mos off but there are at least 4 close open-source competitors.
Woah, SAM3 just reached 1 million downloads on HuggingFace! https://t.co/zcIBDPWPOI
Woah, SAM3 just reached 1 million downloads on HuggingFace! https://t.co/zcIBDPWPOI
"May my heart be kind, my mind fierce, and my spirit brave.” ~K. Forsyth https://t.co/vWMZ7ItNkA
From complex customer feedback, @SurveySparrow is enabling real-time insights. Using robust, compliant AWS infrastructure & genAI solutions, the startup is helping marketers get value from data & scale with confidence. https://t.co/GINfqOZeLD
Experts say that a fraud that has long existed has now become both easier and faster. https://t.co/VPyLolODNd @CSOonline
AI translation is replacing interpreters in GP care here’s why that’s troubling https://t.co/VV71XxDurD @ConversationUS @ConversationUK
AI 'friend' could help you plan your next vacation https://t.co/EqK69XGfiY @FuturityNews
The World is Your Canvas Painting Promptable Events with Reference Images, Trajectories, and Text https://t.co/MAsRMGjV1R
discuss: https://t.co/uefcEJi9cw
nice https://t.co/WbiKCQ8UAg
Starlink is connecting more than 9M active customers with high-speed internet across 155 countries, territories and many other markets. Thank you to all our customers around the world! 🛰️❤️🌎 → https://t.co/lJSdYGR9qN https://t.co/HpnDaKmJyL
Source Control in VS Code now displays Git stashes 📦 Your hidden stashes just became easy to manage! https://t.co/o8TBYbBo7W
“Our wounds are often the openings into the best and most beautiful part of us.” ~ D. Richo https://t.co/X3fflIc0cJ
60 of our biggest AI announcements in 2025 https://t.co/xYbrL2uFlT @google
Yes, AI Is Really Impacting The Job Market. Here's What To Do. https://t.co/1zzOzer5tf @Josh_Bersin
Introducing Vector Lab, a new type of AI image editor Vector Lab exposes control of the underlying features in the diffusion process, allowing you to nudge an image around latent space. https://t.co/Or5t6VbsoB
I partnered with Dreamina to test out their new 3.5 Video model (Seadance 1.5) which now comes with native audio. I created this short scene to put the model through its paces. I was most impressed with the synthetic performance which was more subtle and less uncanny than Veo 3.1. It was easy to prompt consistent voices for my characters. Also, check out the video quality and detail, especially elements like the teeth, which other models struggle with. All sound, apart from the Atmos track and the music, was generated in the model with the video. It's a great model that has already earned its place in my toolkit. Nano-Bananna for the image prompts, Topaz upscale and Suno music. @dreamina_ai
My beloved father has passed away. He fought with extraordinary courage until the very end with grace, strength, and a quiet dignity that defined his life. Even in his final days, when words were gone, his presence spoke louder than anything else ever could. Papa was my anchor, my example, and my definition of resilience. His love, values, and spirit will live on through all of us who were blessed to know him. I will share more when I’m able. For now, I ask for a moment of reflection and remembrance for a remarkable life.
【報告】AIエージェントが最適化プログラミングコンテストにて初優勝 🐟 ブログ：https://t.co/BJut3OjtLA 2025年12月14日に開催された「ALGO ARTIS プログラミングコンテスト2025 師走（AtCoder Heuristic Contest 058）」において、Sakana AIが開発するAIエージェント「ALE-Agent」（AtCoder アカウント名：fishylene）が、804名の参加者の中で1位を獲得しました。 SakanaAIはAtCoder株式会社様のご協力のもとこれまでアルゴリズムエンジニアリングタスクのベンチマークALE-Benchを開発し、独自のALE-Agentを複数のコンテストへ継続的に参加させていただきました。数ヶ月間にわたる改善と試行錯誤を積み重ねてきた中での今回の結果は、数時間単位の高度な最適化タスクにおいても、AIが人間のトップエキスパートに匹敵するパフォーマンスを発揮しうる段階に達したことを示す、一つの重要なマイルストーンだと考えています。 ALE-Agentは、AIが強みとする計算量や思考回数を最大限に活かしつつも、作問者の想定になかった効率的な焼きなまし法の近傍操作を自ら発見するなど、非常に興味深い挙動を示しました。詳しくはブログや下記のログ・技術分析をご覧ください。 ログ・技術分析： https://t.co/Up6rNz2njG 素晴らしい問題を提供いただいた主催の株式会社ALGO ARTIS様、そして継続的にご協力をいただいているAtCoder株式会社様に深く感謝申し上げます。今後も、AIを人間の探索能力を拡張するパートナーとして位置づけ、実社会の複雑な問題解決に向けた研究に邁進してまいります。
【報告】AIエージェントが最適化プログラミングコンテストにて初優勝 🐟 ブログ：https://t.co/BJut3OjtLA 2025年12月14日に開催された「ALGO ARTIS プログラミングコンテスト2025 師走（AtCoder Heuristic Contest 058）」において、Sakana AIが開発するAIエージェント「ALE-Agent」（AtCoder アカウント名：fishylene）が、804名の参加者の中で1位を獲得しました。 SakanaAIはAtCoder株式会社様のご協力のもとこれまでアルゴリズムエンジニアリングタスクのベンチマークALE-Benchを開発し、独自のALE-Agentを複数のコンテストへ継続的に参加させていただきました。数ヶ月間にわたる改善と試行錯誤を積み重ねてきた中での今回の結果は、数時間単位の高度な最適化タスクにおいても、AIが人間のトップエキスパートに匹敵するパフォーマンスを発揮しうる段階に達したことを示す、一つの重要なマイルストーンだと考えています。 ALE-Agentは、AIが強みとする計算量や思考回数を最大限に活かしつつも、作問者の想定になかった効率的な焼きなまし法の近傍操作を自ら発見するなど、非常に興味深い挙動を示しました。詳しくはブログや下記のログ・技術分析をご覧ください。 ログ・技術分析： https://t.co/Up6rNz2njG 素晴らしい問題を提供いただいた主催の株式会社ALGO ARTIS様、そして継続的にご協力をいただいているAtCoder株式会社様に深く感謝申し上げます。今後も、AIを人間の探索能力を拡張するパートナーとして位置づけ、実社会の複雑な問題解決に向けた研究に邁進してまいります。
excited to support this awesome evals work by @SophontAI the largest evaluation suite for assessing medical capabilities of LLMs across 15+ environments is now live on our hub! https://t.co/k59EIwOZzH
We're releasing Medmarks v0.1, the largest completely open-source automated evaluation suite for assessing the medical capabilities of LLMs! Developed in our @MedARC_AI community, w/ support from @PrimeIntellect So far we’ve explored 46 models to figure out the best! https://t.
We're excited to support @SophontAI's Medmarks release The largest completely open-source evaluation suite for medical capabilities Built using verifiers and published on the Environments Hub: https://t.co/hFNFZgz0TR
We're releasing Medmarks v0.1, the largest completely open-source automated evaluation suite for assessing the medical capabilities of LLMs! Developed in our @MedARC_AI community, w/ support from @PrimeIntellect So far we’ve explored 46 models to figure out the best! https://t.
US mortgage lenders insure against artificial intelligence screening errors https://t.co/ePOp7O2ohn @leee_harris @ft
Why AI is a nightmare for the EU https://t.co/1Y8RxBo6eL @calder_mchugh @politicomag @POLITICOEurope
This is huge. First signs of incoming 1000x acceleration of scientific progress. "Researchers found that when scientists use AI, their productivity soared. The biggest jump was in the social sciences and humanities, where output increased by 59.8%, while biology and life sciences saw a 52.9% increase." "Meanwhile, in physics and math, the scientists report a 36.2% boost." "LLM adoption is associated with a large increase in researchers' scientific output," wrote the team.
Introducing TimeFusion, our new multimodal foundation model that unlocks a unified language between humans and sensors. For decades, organizations had access to billions of sensor signals across industry, infrastructure, energy systems, and personal devices but almost none of that data has been easy to understand or act upon directly. TimeFusion changes that. That means you can *ask* a machine about vibration anomalies, *generate* new signals from a text description, or *forecast* what comes next — all in plain English. Here is how it works: 🔹 TimeFusion is a general sensor–language fusion model: a 2-billion-parameter transformer trained to ingest and produce both natural language and raw time-series signals in a single continuous framework. 🔹 Unlike previous approaches that compress sensor data into narrow text-only formats, TimeFusion uses Universal Tokens to combine time-series signals and language inside one shared vocabulary. This enables the model to truly understand physical data instead of translating it through various hacks — and to perform forecasting, anomaly detection, filtering, imputation, captioning, QA, and generation through one unified interface. 🔹 TimeFusion outperforms much larger models like GPT-5, Claude Sonnet, GLM 4.6 and others on sensor-related tasks despite being orders of magnitude smaller. And the model isn’t just translating signals into text. It can already do powerful text-to-signal transformations: forecasting the future of a waveform, reconstructing missing data, filtering noise, or reshaping a signal based on a natural-language prompt producing new signals rather then text as output. This opens the door to an entirely new category of interfaces with the physical world — where engineers, operators, doctors, city systems, and even consumers can converse with the machines and environments around them, instead of digging through raw numbers and graphs. A new way to talk to the physical world is here 🌍 #PhysicalAI
Haha got roasted by ChatGPT https://t.co/GPxVThtBjb
Stop repeating your instructions. 🛑 Copilot now supports Agent Skills, the open standard by @AnthropicAI. Package your specialized expertise into skills once and use them everywhere. Try it in: ✅ agent mode in @code Insiders ✅ Copilot coding agent ✅ Copilot CLI Here's how to get started. 👇 https://t.co/AY0EXNanwC
2025 was a busy year. From model launches to Labs experiments to scientific breakthroughs, it can be tough to keep up with all of the new ways AI can help make your life easier. So, here are 3 of our favorite helpful AI tips (and you can find 37 more in the link below): 1. Make an interactive simulation with Gemini 3: If you're researching something complicated like mortgage loans, Gemini 3 in AI Mode can make you a custom-built interactive loan calculator, so you can compare two different options and see which offers the most long-term savings. 2. Find your style with a selfie: Upload a selfie into our virtual try on tool. Nano Banana will generate a full body digital version of you to virtually try on the latest styles. 3. Get homework help with Guided Learning: Guided Learning in the @GeminiApp an interactive study partner that helps you build a deeper understanding of any topic. It can generate study guides from uploaded course material, walk you through debugging problematic code or simply explain topics with helpful videos and images. Check out the rest, and share your own tips with us! ⬇️ https://t.co/o94gbstesR
I wrote a post for @Variety about the complexities of #Native identity. It’s not an easy subject & would benefit from more listening and understanding. #sacheenlittlefeather #NativeTwitter https://t.co/5PsUDHKVrE
【好評受注中】 『魔太郎』さんイラスト 「SAMURAI -零-」 原型制作：かふき https://t.co/ot6bED7Akr #ホットビーナス #エロホビ https://t.co/pn91oWUrNW
