AI with Ravi

Thursday Noon Edition

March 26, 2026
25+
Tweets
7
Topics
8h
Window
Gemini 3.1 Flash Live Drops
2 items
Google launches Gemini 3.1 Flash Live — its biggest upgrade to real-time voice AI — with 2x longer context, noise filtering, and dynamic tone matching.
Josh Woodward announced the model powers Gemini Live on both Android and iOS, calling it faster, smarter, and more emotionally intelligent. Google is betting that a mid-tier model with excellent latency beats a frontier model that stutters — positioning Gemini Live as a genuine daily-driver voice assistant. Logan Kilpatrick confirmed the team spent over a year on model + infrastructure improvements, calling the result a "step function."
@joshwoodward · 19m · 179 likes, 11K views — reposted by @GeminiApp | Also: @OfficialLoganK · 56m
Early-access tester shows Gemini 3.1 Flash Live leading all competitors on the Big Bench Audio benchmark at 97% speech recognition accuracy.
Ashutosh Shrivastava shared benchmark charts showing Flash Live significantly outperforming OpenAI's realtime models, Gemini 2.0, and Nova 2.0 on speech recognition tasks. The noise-filtering improvement is the standout practical gain — working in real-world noisy environments, not just lab conditions.
@ai_for_success · 49m · Early access review
The Vibe Coding Reckoning
5 items
Andrej Karpathy says the hardest part of building MenuGen wasn't the code — it was the DevOps: payments, auth, database, security, domain names, "like IKEA furniture."
Patrick Collison quote-tweeted Karpathy's original blog post about vibe coding being exhilarating locally but a slog to deploy. AI can now write application logic fluently, but the unglamorous plumbing of real deployment remains stubbornly human. This is the new bottleneck — and the next frontier for AI tooling to crack.
@karpathy · 4m · quoting @patrickc · 104 likes, 6.9K views
Carlos E. Perez argues Karpathy's No Priors interview is more radical than summaries suggest: "Code's not even the right verb anymore."
He's not saying AI is a better tool — he's saying the structure of knowledge work itself is changing. When someone at Karpathy's level reframes the vocabulary, it usually means the old mental model has already broken.
@IntuitMachine · 1h · 11 likes, 1.2K views
Ethan Mollick revisits a year-old prediction that AI would write 100% of code — and notes Claude Code now writes a remarkable percentage. His sharper point: adoption is the barrier, not technology.
The original prediction by @kimmonismus from March 2025 said AI would write 90% of code within 3-6 months and 100% within a year, citing Dario Amodei. Mollick's reframe — that the tools are ready but organizations aren't — connects to the Jevons paradox framing that Chollet and Levie were exploring in parallel.
@emollick · 3h · 241 likes, 24K views
François Chollet predicts more software, more demand for software engineers, and a lot more token consumption — Jevons paradox in real time.
Aaron Levie's original observation: companies outside tech are realizing they can now afford software projects AI makes feasible. Chollet adds the supply-side corollary: when software gets cheaper to write, the world wants infinitely more of it. Both sides of the equation point to growth, not displacement.
@fchollet · 2h · quoting @levie · 208 likes, 17K views
An immunologist with no CS training launched a complex science app with Codex overnight and "just froze, staring at the screen" in disbelief.
Derya Unutmaz, MD was quote-tweeting JB's Uncivilized.fun — a Civilization clone with natural-language AI diplomacy built in 48 hours. Domain experts are shipping real software with conversational AI, and the gap between "toy demo" and "actual product" is collapsing fast.
@DeryaTR_ · 3h · quoting @jamie247 · 70 likes, 7K views
Five separate voices in 8 hours — Karpathy, Chollet, Mollick, Perez, and Unutmaz — all independently converged on the same conclusion from different angles. The coding paradigm shift isn't a prediction anymore. It's a Wednesday.
ARC-AGI-3 & the Benchmark Wars
2 items
12 hours into the ARC Prize 2026, the human high score on ARC-AGI-3 is still below the template baseline — with Stochastic Goose leading at 0.25.
Greg Kamradt flagged the leaderboard as worth watching. François Chollet's new benchmark was designed to crush current AI models, and early signs suggest it's working — the scores indicate ARC-AGI-3 measures something frontier models can't yet brute-force. The competition is just getting started.
@GregKamradt · 9h · reposted by @bag_of_ideas
A research paper on arXiv had 2 citations after 11 months — the same work as an accessible blog post hit 12 million views in one day.
Jia-Bin Huang highlighted the power of medium, referencing Google Research's TurboQuant — a compression algorithm that reduces LLM KV-cache memory by 6x with zero accuracy loss. The science was always there; the blog post just let people find it.
@jbhuang0604 · 14h · reposted by Christian Langreiter
AI Infrastructure & Capital Flows
2 items
A Morgan Stanley diagram mapping OpenAI's ecosystem capital flows went viral at 257K views — showing revenue and investment cascading through Microsoft, Oracle, Coreweave, Amazon, and Nvidia.
Lydia DePillis called it "hell of a graphic." The diagram makes visible what's usually abstract: OpenAI sits at the center of a financial web where compute dollars flow down through cloud providers to chip makers to data center lessors. It's the AI economy in one picture.
@lydiadepillis · 20h · reposted by @fastml_extra · 4.1K likes, 257K views
Cloudflare discovered a one-line Kubernetes config change that reduced restart time from 30 minutes to 30 seconds, saving 600 hours a year.
The fix — adjusting fsGroupChangePolicy — is absurdly simple, which is exactly the point: infrastructure complexity is now so deep that million-dollar savings hide in single-line changes nobody thinks to look for. As Karthi noted: "K8s is modern world Cobol."
@Cloudflare · 3h · 64 likes, 16K views | @KarthiMahade · 2h
Local AI & Privacy-First Tools
1 item
Niels Rogge demonstrates running Nvidia's NemoClaw with Qwen3.5-27B entirely locally via Telegram — no API costs, no data leaving the device.
NemoClaw is a security sandbox built on OpenClaw that restricts file and network access for the AI agent. The video shows a complete agentic assistant on consumer hardware through a chat interface people already use. Open-weight models are now good enough that the local-vs-cloud tradeoff has genuinely tipped for many use cases.
@NielsRogge · 2h · reposted by AK · 185 likes, 20K views
Brain-Computer Interface Milestone
1 item
Neuralink's ALS patient Kenneth Shock progressed from speaking aloud to silently thinking words decoded into speech — using ML to map neural activity to phonemes.
Kenneth received the N1 implant in January 2026 and is part of Neuralink's VOICE clinical trial. The key advance is phoneme-level neural decoding — learning the smallest units of sound rather than whole words — making the system far more generalizable. This is thought-to-speech crossing from research demo to clinical reality.
@WesRoth · 3h · quoting @neuralink · 17 likes, 1.7K views
The Lighter Side & Quotables
3 items
An AI detector flags Abraham Lincoln's Gettysburg Address as 96.2% AI-generated — a perfect illustration of why these tools remain unreliable.
ZeroGPT's false positive on one of history's most famous speeches proves these tools measure "sounds like GPT" rather than "was made by GPT." Well-crafted human prose triggers the same statistical patterns.
@rohanpaul_ai · Mar 25 · reposted · 74 likes, 26K views
Richard Dawkins turns 85 on #ScienceAppreciationDay: "Without science, you probably wouldn't have reached your latest birthday."
A characteristically crisp framing on a day that shares his birthday. Science saves lives, he notes, and it's also humanity's best shot at universal understanding.
@RichardDawkins · 33m · 184 likes, 7.4K views
The US lost the 5G race because it didn't give one Turkish mathematician a green card — and 100K people felt the sting.
Erdal Arikan solved a fundamental problem in information theory at Caltech/MIT, couldn't find a US appointment, returned to Turkey — and Huawei used his discovery to build one-third of the world's 5G infrastructure. Today Huawei holds over two-thirds of related patents.
@AlecStapp · 3h · reposted by @garrytan · 2.9K likes, 100K views