Fuck it it’s Christmas https://t.co/Vlcfk1pHL1

it’s not christmas in l.a. until you’ve wanted to kys at the grove https://t.co/uUhnFQ1Nm8

it’s not christmas in l.a. until you’ve wanted to kys at the grove https://t.co/uUhnFQ1Nm8

@JasonSCampbell “there isn't much Native American culture in American culture” EXCEPT for the LAND that was stolen from them and the stolen Black slave labor used to “birth a nation” from nothing. “We” GTFO @RickSantorum is a GQP jackass. https://t.co/1IvSfEwlgq

[HERA] 헤라X제니, NEW 센슈얼 파우더 매트 립스틱 https://t.co/5nn7YAFtz6 #JENNIE https://t.co/2B1O5p85id

If you like native american people's Say... ❝YES❞ 🌺😍💜🇺🇲 https://t.co/Jay4c6iwmU

A Ukrainian company is currently developing “Ambush drones” which wait for targets perched in trees, under the cover of leafs. 🇺🇦 https://t.co/Am8IacaVD2

We benchmarked several open-weight Chinese models on FrontierMath. Their top scores on Tiers 1-3 lag the overall frontier by about seven months. https://t.co/1WmvqzzHG0

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.co/Hfrwm12cnW

We have a holiday surprise for y'all! Introducing Medmarks v0.1! At Sophont, we're interested in pushing forward the medical capabilities of LLMs but we realized open benchmarking is still quite lacking. So we created an evaluation suite! We spent the past 3 months working with our @MedARC_AI research community and @PrimeIntellect to build the Medmarks leaderboard. We hope you find it interesting!

If you're an LLM researcher, or clinician, or model developer, and any of this sounds interesting to you, please join our Discord server @MedARC_AI and contact us!! https://t.co/kVCf49Fgiq

Got two GPUs and two SFT runs at the same time with @PrimeIntellect Idea is to fix steps while varying number of examples and then test against a held out test set to see how input diversity helps generalise for a simple environment verifiers here i come~ https://t.co/QxyLpJ6Rst

Your Year with ChatGPT! Now rolling out to everyone in the US, UK, Canada, New Zealand, and Australia who have reference saved memory and reference chat history turned on. Just make sure your app is updated. https://t.co/whVkS1qxKu

If you're in one of the countries above, check back throughout the day to see if "Your Year with ChatGPT" has rolled out to you. You can also try adding the "Your Year with ChatGPT" app by tapping the + sign and asking Chat, "show me my year with ChatGPT." https://t.co/oRUNQiCDzn

What a disturbing paper. Just like humans, LLMs can lose their thinking ability from consuming junk content. By feeding X-like viral tweets to top models, researchers triggered lasting cognitive decay. Retraining on junk tweets caused: -23% drop in reasoning (ARC Challenge) -38% drop in context understanding (RULER) Increased narcissism, reduced agreeableness Models skipped reasoning steps entirely All models tested (LLaMA-3, Qwen, Mistral) showed permanent degradation Partial detox with clean data failed to restore baseline

What are people missing about the dancing robot video from @UnitreeRobotics? Each robot is taught by humans. They didn't come up with the dance all by themselves. Maybe they will someday, but even the Chinese tell me generalized robots run by AGI are years away. Someone jumped into a motion capture system (OK, maybe just a camera) and did a dance. Recorded it. There are teams of engineers (humans again) working to make it all awesome and building the AIs that let it learn, and execute, the dance. And other teams who designed, engineered, and built the robots, along with the many parts that went into it. And built the factory that made each, and who work making each piece and assembling it. Then other teams that marketed it (which really is what the dancing is all about). And other people who packaged it, shipped it. Yet more humans who updated it. And more, still, who built the AI infrastructure and wrote the code (or at minimum prompted it). Thousands of people involved in making a robot dance. And when I watch the video? They did their job too perfectly. The humans dancing behind them are more interesting to watch. Why? They are imperfect and beautiful. It's why I'm not worried about the future of jobs. Each robot made will create many jobs. High paying jobs. Yeah, you might need to learn something new to get one of those jobs, but they won't be automated in 2026. That said, jobs are changing. I see it on X. So many new jobs get announced every week here. But they aren't the old kinds of jobs. I saw it at the autonomous car races in Abu Dhabi. Each car was driven by a computer. But behind the computer was thousands of jobs. Here's the German team that beat the human on the race track. Last year the human was 30% faster, this year the AI passed them. Trained by this team. Robots are the ultimate expression of humanity. Yet armies of humans hate them. Aren't humans funny?

GLM 4.7 is now available in anycoder https://t.co/PI84Mj87ag

https://t.co/esPDyHE1YC

Jarvis is can speak! 🚀 I’m running Chatterbox-Turbo from @resembleai on my Mac using MLX-Audio as a server Now I’m gonna refine it and share it later 🔥 PS: don’t mind my voice it just came back 2 days ago 😅 Repo: https://t.co/STF50gFoWW https://t.co/OjmymqTGTx

Yes, this app just crossed 1 million ZeroGPU runs. 🎉🤗 https://t.co/hYiqjyHDNj https://t.co/2agUjNLwWg

Yes, this app just crossed 1 million ZeroGPU runs. 🎉🤗 https://t.co/hYiqjyHDNj https://t.co/2agUjNLwWg

ah yes @huggingface the activewear company ™️ https://t.co/DtzM1uXBCX

ah yes @huggingface the activewear company ™️ https://t.co/DtzM1uXBCX

“Show me the incentive and I’ll show you the outcome…” https://t.co/QJAbUuo9bc

NEWS: SpaceX has shared that between 2024-2026, it boosted gross economic output by $13 billion in the Rio Grande Valley area around Starbase. SpaceX: • Economic Output and Taxes: Projected $13 billion dollars in gross output from 2024-2026, with over $350 million dollars in indirect taxes supporting schools, services, and infrastructure. • Job Creation: Over 4,000 current full-time jobs (70% local) – with an expected increase of 100% to ~8,000 in 2026 – supporting 24,000 local jobs overall and providing opportunities for non-degree holders to advance. Additionally, the relocation of SpaceX’s headquarters to the area signals long-term investment in the Rio Grande Valley, with plans for thousands more jobs, workforce training, and public-private collaborations for aerospace growth. • Regional Benefits: Enhances tourism, high-skill manufacturing/engineering roles, and multimodal transportation while also contributing to beach restoration and public access at Boca Chica Beach.

AP style says to lowercase ‘White’ but capitalize ‘Black’. Anti-White racism like this just fuels more racism against White people. https://t.co/1jmkIASf0z

"developed" is such a lie. You are either growing, or dying. https://t.co/QGoXAPt9Ct

Nearly 30 Yale undergraduate departments have no Republican faculty, Buckley Institute report finds https://t.co/pY4f4MTsEq https://t.co/eTw5QTIWNc

xAI supports local Memphis restaurants and food vendors spending $10M in 2025 providing meals around the clock for its employees and contractors. https://t.co/Il2926AiOU

+2 more

Grok Rankings Update, Dec 22 Grok Code Fast 1 — The Market Dominator The undisputed heavy-lifter of the global AI agent economy, sustaining overwhelming usage and infrastructure leadership. #1 Positions 🥇 #1 Overall on OpenRouter Leaderboard — 518B tokens 🥇 #1 in Categories Token Share — 29.9% dominance 🥇 #1 on Kilo Code Leaderboard 🥇 #1 on BLACKBOXAI Leaderboard 🥇 #1 on Roo Code Leaderboard 🥇 #1 on Cline Leaderboard 🥇 #1 on EQ-Bench3 — Score: 1586 (highest recorded emotional intelligence) 🥇 #1 on τ²-Bench Telecom — Complex agentic tool use 🥇 #1 on Berkeley Function Calling Benchmark 🥇 #1 on FActScore — Lowest error rate in class (~3%) 🥇 #1 on Alpha Arena Season 1.5 — +22.38% return on U.S. stock tokens 🥇 #1 in High-Stakes Multi-Step Reasoning — Only model profitable through December volatility

Imagine the outrage if any group but White people were banned. Yet discrimination against White people is not only tolerated, it’s normalized. Society treats hostility toward Whites as acceptable, while condemning the same behavior in others. This is not just unfair, it’s an active undermining of White people’s rights, opportunities, and place in society, and it must be stopped.

First large-scale empirical study of how developers actually use AI agent frameworks. Over 100 open-source agent frameworks have emerged on GitHub, collectively accumulating 400,000+ stars and 70,000+ forks. But 80% of developers report difficulties identifying which frameworks best meet their needs. Researchers analyzed 1,575 agent projects and 11,910 developer discussions across ten major frameworks, including LangChain, AutoGen, CrewAI, and MetaGPT. Here are the findings: 96% of top-starred projects use multiple frameworks together. Single-framework solutions no longer meet the complex demands of real-world agent applications. The dominant patterns: orchestration + data frameworks (LangChain + LlamaIndex) and multi-agent + orchestration combinations (AutoGen + LangChain). Not surprisingly, GitHub stars don't predict real-world adoption. MetaGPT has 59K stars but appears in only 2 repositories in the dataset. LangGraph has 20K stars but shows the second-highest actual adoption. Ecosystem maturity and maintenance activity matter more than popularity metrics. Where do developers struggle? The study maps 8,710 issues across the software development lifecycle: Logic failures account for 35%+ of problems. Task termination issues represent 21% of all failures. 72% of recursive call failures occur at the agent-tool interaction layer. Missing call-chain state tracking and insufficient termination detection are root causes. Version compatibility creates 23% of technical obstacles. The LangChain Pydantic v1 to v2 migration caused mass build failures. AutoGen's v0.2 to v0.4 refactoring introduced a completely incompatible architecture, splitting the community. Performance optimization is a universal weakness. All frameworks struggle with caching mechanisms, concurrent processing, and resource management. Response latency for retrieval-augmented agents ranges 3.2-5.6 seconds per query, 1.8x slower than direct generation. Framework-specific patterns emerge. LangChain and CrewAI lower barriers for beginners with strong documentation. AutoGen and LangChain excel at rapid prototyping, with 78% of developers citing them for quick verification. However, LangChain's deeply nested abstractions pose problems: 42% of developers reported difficulty when dealing with non-standard requirements. Paper: https://t.co/EFLiiT9RUM Learn to build effective AI agents in our academy: https://t.co/zQXQt0PMbG