Overview
AI’s battle for the stack heated up, from Google’s push to run PyTorch on TPUs to xAI’s model drumbeat and DeepMind’s safety tools. Builders showed what agents can already do in the browser and in no-code app creation. Out in the real world, robots hit the stage, drones pushed past jamming, and space made a splash with an interstellar visitor. The culture side ran from Cybertruck hype to CS2 market swings, while policy took the mic on traditional medicine.
The big picture
Google’s TorchTPU leak aims at CUDA’s moat
A leaked project suggests Google is making PyTorch run natively on TPUs, trimming the friction of moving off Nvidia’s CUDA-centric world. With PyTorch dominant in research, this could widen hardware options if execution lands on time and gains support. 🔗 Post link
DeepMind rolls out Gemma Scope 2 for interpretability
Google DeepMind introduced tools to inspect how Gemma 3 models represent concepts, using sparse autoencoders and transcoders across sizes. It ships with code, data and demos to help researchers trace internal behaviour and spot risks. 🔗 Post link
Grok 4.20 teased, with speed and swagger
Mario Nawfal says Grok 4.20 lands soon, promising faster responses, stronger multimodal skills and API upgrades, with benchmark bravado to match. 🔗 Post link
Elon Musk touts Grok 4 Heavy’s study-group agents
Musk calls Grok Heavy the smartest model, highlighting a multi-agent setup that compares outputs before returning a final answer, aimed at tougher tasks. 🔗 Post link
A viral emergency test crowns Grok “the helper”
A side-by-side video shows Grok opting to run a red light if safe to save a life, while others refuse and cite the law, sparking debate on AI risk and priorities. 🔗 Post link
“I built an app by talking” hits the feed
Markandey Sharma’s thread shows Mocha turning a spoken idea into a working app with database, backend and hosting, plus extras like payments and AI image tools. Users praise speed, while noting crashes and high credit use. 🔗 Post link
A markdown file beats a lab, says Sawyer Hood
Hood’s open dev-browser, which lets models run code to control the browser, outpaced Claude Code’s Chrome extension on a real task, finishing faster, cheaper and cleaner. 🔗 Post link
Robots take the stage in Chengdu
Unitree’s G1 humanoids danced and threw Webster flips at Wang Leehom’s concert, showing off balance and coordination from modern control and learning methods. 🔗 Post link
Russia touts a 65 km fibre FPV coil
A claimed demo shows a drone trailing a fibre spool over mountainous terrain without signal loss, pointing to jam-resistant long-range strikes and new counters. 🔗 Post link
Google and NASA explain interstellar comet 3I/ATLAS
Google teamed with NASA to answer public questions on 3I/ATLAS during its closest pass, tackling myths and showing the science behind its behaviour and jets. 🔗 Post link
Space memes recap a messy year of launches
A fast-cut “brainrot” video skewers delays and explosions across the industry, from Starship to Vulcan, mixing humour with real frustrations. 🔗 Post link
Cybertruck test-drive rallying cry
Tesla Owners Silicon Valley amplifies Musk’s pitch to get behind the wheel, with a montage of rough terrain shots and chatter around SpaceX bulk buys and sales targets. 🔗 Post link
CS2 skin market has a moment
An Interstellar meme channels regret over cheap glove cases after rare drops were reportedly disabled, stirring trader nostalgia and speculation on prices. 🔗 Post link
Modi urges healthcare to think long term
India’s PM calls for modern science to sit alongside traditional medicine, arguing for evidence, trust and resilience to serve future generations. 🔗 Post link
Why it matters
AI infrastructure is in flux. If PyTorch runs on TPUs without rewrites, hyperscalers get pricing power and supply options. Tooling and community support will decide how much real workload moves. At the same time, interpretability kits like Gemma Scope 2 help researchers see what models represent, which matters as systems grow and regulators ask harder questions.
Model rivalry is becoming about style as much as scores. Grok’s push mixes speed claims, multi-agent orchestration and a public stance on values, while competitors focus on caution and compliance. Expect product choices to hinge on policy settings by sector, not just benchmarks.
Agents are getting practical. Mocha and dev-browser show that end-to-end app creation and code-driven web control can already save time, but reliability, cost and guardrails are the sticking points. The lesson is simple, give models tools and a tight loop, then judge on outcomes, not hype.
Robots and drones are crossing from labs to live use. Stage shows sell social acceptance, while fibre-linked FPV points to tactics that ignore jamming. Both raise safety questions, from public venues to battlefields, and will invite new standards.
Science outreach still matters. Clear, engaging explanations around 3I/ATLAS cut through rumours and build trust. Memes about rockets remind us attention is scarce, so good information needs good packaging.
Consumer excitement and markets move fast. Cybertruck clips keep the fanbase engaged even as targets bite. In CS2, a small rules change can swing digital asset values overnight, a reminder that platform policy is market policy.
Health policy’s long view is back on stage. Blending evidence with tradition is hard to govern, but if done well it can widen access and resilience without abandoning science. That theme runs through the day, from AI memory debates to medicine, ask what to keep, what to test, and what to build next.
