Google DeepMind dropped Gemma 4 in four sizes under Apache 2.0 — and the 31B dense model already ranks #3 among all open models on Arena AI.
François Chollet announced during a live Keras community call. Four sizes (E2B, E4B, 26B MoE, 31B Dense) with native function-calling, structured JSON, 128K-256K context. Apache 2.0 license means no MAU caps, no use restrictions — Google betting openness wins the developer war against Meta's Llama.
Demis Hassabis posted benchmarks showing Gemma 4 outperforming models over 10x its size — noting the x-axis is log scale.
Gemma 4 31B scores 89.2% on AIME 2026, 84.3% on GPQA Diamond, 80.0% on LiveCodeBench v6 — territory previously reserved for closed frontier models. Architecture from Gemini 3 research cascading down to open weights.
Omar Sanseviero published a visual architectural deep dive of Gemma 4 — from per-layer embeddings to vision and audio encoders.
Community documentation that accelerates adoption faster than official blogs. Traces full architecture from input embeddings through MoE routing to output heads.
NVIDIA confirmed Gemma 4 31B runs 2.7x faster on RTX GPUs using llama.cpp — joint optimization with Georgi Gerganov.
Consumer-GPU-ready on day one. Hardware-software co-optimization between NVIDIA's AI PC team and llama.cpp maintainer that used to take months.
Anthropic is testing "Conway," an always-on persistent agent that transforms Claude from chatbot to autonomous digital twin.
Leaked project reveals dedicated web workspace running Claude Code continuously with webhooks, Chrome interface, notifications. A CNW ZIP extension standard would let developers build custom tools and UI tabs — essentially an app store for persistent agents.
Andrew Curran: if OpenAI and Anthropic both finished training capable models in early March, the convergence may be purely a result of scale.
Two labs hitting similar thresholds simultaneously suggests performance is downstream of compute budgets, not unique architecture. Implication: capability becoming commodity, differentiation shifts to product and ecosystem.
Ethan Mollick declared the RAG era "short-lived, but intense" — no longer the dominant paradigm for context.
Long context windows, tool use, and persistent memory displacing chunked retrieval. RAG not dead but absorbed into larger toolkit rather than remaining central organizing principle.
François Fleuret asked Claude to diagnose a model checkpoint and called the result "baffling" — strongly recommended.
The transformer-diffusion hybrid researcher simply prompted Claude to inspect weights and flag problems. A sign LLMs are becoming genuine research collaborators, not just code assistants.
Maithra Raghu predicted more capable agents will require more humans, not fewer.
Long-horizon AI creates more effort at handoff points, and stakes of autonomous decisions rise with capability. Counter to replacement narrative but aligned with actual enterprise deployments.
Pieter Levels admitted vibe coding in production is "very dangerous" — cutting off all database access for AI-generated code.
The indie developer who hit $100K MRR with a flight sim built in 3 hours is now sandboxing AI code at infrastructure level. Driven by 2026 wave of exposed databases and tokens from AI-coded apps. His response may become template for the vibe-coding community.
Marc Andreessen: "AI increases workload. Many such cases."
AI tools generate output so fast that reviewing, debugging, and integrating creates net new work. Echoes the growing "AI productivity paradox" where adoption increases throughput but also cognitive load of quality control.
Garry Tan reposted Elvis Sun's argument that "AI slop" is the future of software engineering — mass code review is legacy thinking.
FastCompany piece frames it as velocity versus craft. Strong reactions because it challenges the professional identity of developers who prize clean code. Central culture war in software: AI slop at scale vs AI-assisted craftsmanship.
Simon Willison warned that the Axios npm attack started with "very sophisticated social engineering" targeting a single developer.
North Korean UNC1069 cloned a company founder's identity, built a branded Slack workspace to steal npm credentials. Two trojanized Axios versions (100M+ weekly downloads) shipped a cross-platform RAT. Entire exposure window: roughly three hours.
Netflix open-sourced VOID — a video model that removes objects along with all physical interactions they induce.
Remove a person holding a guitar, and the guitar falls naturally. Novel quadmask conditioning on CogVideoX, trained on Blender physics simulations. On HuggingFace under open license. More tech companies doing open source is "the good timeline."
300+ humanoid robots from 26 brands will compete in Beijing's half marathon on April 19 — full 21km alongside humans.
Scaled 5x from last year: 76 institutions, 80+ corporate teams, 20 university teams. Robots over 75cm must complete full distance continuously. 38% capable of fully autonomous navigation. New obstacle challenge simulates disaster recovery.
Fabian Gloeckle translated an entire graduate math textbook into Lean using 30,000 LLM agents.
Open-source algebraic combinatorics project. Large-scale multi-agent inference that actually works. Formal math verification bottlenecked by manual labor for decades — brute-force agent scaling may have crossed the quality threshold.
Fei-Fei Li kicked off 11th year of CS231n at Stanford — departments represented keep changing.
The original computer vision course evolved into a cross-disciplinary magnet drawing students far beyond CS. Annual ritual serves as quiet barometer of AI education diffusing across the academy.
Heated debate over who best portrayed Lord Rama in Telugu cinema — PVR Narasimha Rao: "NOBODY portrayed Rama and Krishna better than NTR."
Responding to claims that Harinath Raju and Shobhan Babu surpassed NTR, Rao followed up with Rama vs Ravana comparison, calling it an era when "writers, directors and actors knew what they were doing." A reminder the feed carries deep cultural memory alongside its AI signal.
A rare six-faced Bhagwan Kartikeya vigraha discovered in Himachal Pradesh drew 1,000 likes from a feed that typically celebrates neural networks.
Crossover engagement — heritage content resonating with a tech-oriented audience — suggests a community that values tradition as deeply as hyperparameters.
Vishal Misra shared a production thread on the Ramayana film, praising "the obsession and research Nitesh has put in."
In a week dominated by AI imagery discourse, a post celebrating years-long human craftsmanship in filmmaking struck a chord. The slow work of art against the fast work of algorithms.
Today's feed is bookended by two kinds of craftsmanship — the race to build and deploy AI systems at unprecedented speed and the slow, reverent work of preserving and creating culture. The tension between velocity and craft is the feed's recurring bass note.