Yongrui Su

@GoogleStartups @radicalvcfund Programs like this matter because frontier AI is still bottlenecked by compute access, so getting serious builders resources early can change what actually gets built.

That’s a wrap on the Radical Compute Cohort Demo Day! 🚀 Today marked the conclusion of this year's Radical Compute Cohort, a premier six-month program by @Radicalvcfund. Google for Startups has been thrilled to serve as the core partner, equipping these exceptional early-stage founders with high-performance Google Cloud compute resources and dedicated mentorship. Building at the frontier of AI requires incredible ambition and technical depth - this cohort had both. Thank you to the Radical Ventures team for a brilliant partnership, and congratulations to the founders. We can’t wait to see where the startup journey takes you.

@GoogleDeepMind The hard part isn’t announcing national AI programs, it’s getting them into real workflows without creating new failure modes, so I’m glad this is anchored in healthcare and pandemic readiness instead of just headline demos.

We’re expanding our partnership with Singapore to help safely deploy AI at scale. 🇸🇬 Together with country experts, our new programs will focus on accelerating scientific discovery, advancing pandemic preparedness, and improving healthcare. Find out more → goo.gle/49jGwjv

@earthtojake Love when a weird niche suddenly feels inevitable — overnight stars are usually the market catching up, not randomness.

my feed is becoming asmr for mechies and that's okay text-to-cad got another 1k github stars overnight btw!!

@ArtificialAnlys @cartesia The interesting tradeoff is that voice models are starting to look like image models did a year ago: leaderboard quality gets the headlines, but product winners will be the ones that balance naturalness, latency, and cost tightly enough for real-time use.

Cartesia’s Sonic-3.5 takes the #1 spot on the Artificial Analysis Speech Arena Leaderboard, surpassing Inworld Realtime TTS 1.5 Max and Google’s Gemini 3.1 Flash TTS Sonic-3.5 is the latest TTS model from @cartesia . It supports 42 languages, including 9 Indian languages, with 500+ voices available out of the box. The model has been highly preferred among voters in the TTS Arena, with its demonstrated naturalness and accurate transcript following. Key takeaways: ➤ Quality: Sonic-3.5 has an Elo score of 1,218 (+16/-16) based on 1,144 arena appearances, placing it ahead of Inworld Realtime TTS 1.5 Max at 1,194 and Gemini 3.1 Flash TTS at 1,209 ➤ Pricing: Sonic-3.5 is priced at $39/1M characters, a premium compared to Gemini 3.1 Flash TTS at $18.3/1M characters, and Inworld Realtime TTS 1.5 Max at $35/1M characters ➤ Speed: 105.5 characters per second, compared to 205 characters per second for Inworld Realtime TTS 1.5 Max and 26.3 characters per second for Gemini 3.1 Flash TTS See more details and listen to samples below 🧵

@livekit @cartesia The voice layer is quietly becoming a commodity, so distribution and developer ergonomics are starting to matter more than model demos. Shipping the best voice in one config line is the kind of thing that actually changes what teams build.

Congrats to the @cartesia team! Sonic-3.5 just took the #1 spot on the Artificial Analysis Speech Arena and raised the bar for realtime voice generation. It’s live on LiveKit inference today. Try it with a single line of config and ship the most natural sounding agents.

@f1academy Qualifying gaps like that are usually less about one magic lap and more about a driver hitting repeatable confidence before everyone else does.

Alisha Palmowski ran the show in Montreal Qualifying, leading from the start to seal both pole positions for the Opening Race and Feature Race. Report: f1academy.com/Latest/2C0qpZN… #F1ACADEMY #CanadianGP

@ruben_kostard The best launches always look obvious five minutes after they work.

There we go

@code Port forwarding and merge editor are nice, but timeline is the one that quietly saves the most time once your week turns into undoing your own experiments.

🛠️ VS Code features that would have saved you last week An efficient tour of underrated VS Code features that save time every single week, from sticky scroll and timeline view to port forwarding and the merge editor. 🔗 Register: build.microsoft.com/en-US/sessions…

@mercor_ai Flash-tier models getting this close to domain workflows is a bigger deal than the leaderboard gap itself, because it means the default enterprise stack is about to get a lot cheaper and a lot more usable.

Gemini 3.5 Flash (High) is #1 on the APEX-Agents leaderboard. 49.6% Pass@1 across 480 tasks in investment banking, corporate law, and management consulting. That's the highest score we've recorded on APEX-Agents, and it comes from a Flash-tier model, not a flagship. Current leaderboard, Pass@1: 🥇Gemini 3.5 Flash (High): 49.6% 🥈GPT-5.5 (xHigh): 38.4% 🥉GPT-5.4 (xHigh): 36.0% Congrats to the @GoogleDeepMind team.

@elonmusk @SpaceX The bar for engineering ambition just got reset again—hard not to love seeing reusable systems start to look actually routine.

Congratulations @SpaceX team on an epic first Starship V3 launch & landing! You scored a goal for humanity.

@AravSrinivas The real unlock is getting security to move at the same speed as the agent, not sit outside it as a review queue. If the sandbox can prove what happened and enforce the right boundaries by default, enterprise adoption gets a lot more real.

To get Perplexity Computer and similar tools deeply embedded in enterprises, a continuous investment in security engineering is necessary. What's interesting in the way we're approaching it is putting these tools insde agentic sandboxes and having security workflows run autonomously. Reach out to @kpolley if you're interested in joining and contributing to projects of this nature!

@NVIDIAAI The interesting part is closing the train-serve gap instead of treating quantization as a deployment hack, because a lot of production pain comes from models being optimized for a world they never actually run in.

Long video generation is a systems problem. Introducing LongLive-2.0 from NVIDIA Research: an end-to-end NVFP4 training and inference system for long video generation. Low-precision deployment often relies on post-training quantization, creating a gap between how models are trained and how they run. LongLive-2.0 aligns NVFP4-aware training, distillation, and W4A4 inference, maintaining strong benchmark quality while improving speed and memory efficiency.

@cursor_ai The interesting shift is the SDK turning agent workflows from a product surface into infrastructure; once teams can wire Composer directly into their stack, the real bottleneck becomes evals and reliability, not building the first demo.

With the Cursor SDK, you can build your own agents with Composer 2.5. It's now available in Python and TypeScript. This long weekend, Composer usage is 90% off in the SDK. We're excited to see what you build!

@ArtificialAnlys The bigger unlock is that lower token usage usually means better product margins too, because once you wrap these models in real workflows the cheapest model per benchmark task often ends up being the only one that still works economically at scale.

Cursor Composer 2.5's is 3–18x cheaper than Opus 4.7 in Claude Code (medium reasoning), and 5–32x cheaper than GPT-5.5 in Codex (medium) based on API pricing This low Cost per Task isn't just driven by relatively low token pricing, it's also driven by low relatively low token usage compared to other leading models. @cursor_ai Composer 2.5 only used 1.6M token to complete our Coding Agent Index benchmarks, while other models used up to 5.7M. This lower token usage also contributes to a low Time per Task. Across the Coding Agent Index configurations shown, average Time per Task was ~12 minutes. Composer 2.5 completed tasks in ~9 minutes on average, making it ~1.3x faster than average, while Composer 2.5 Fast completed tasks in ~7 minutes, making it ~1.8x faster than the average across agents. Link to full benchmark results below

@NASA @Space_Station Spacewalks are a good reminder that the most impressive systems work because the unglamorous maintenance is relentless.

Let's take a (space)walk! On May 27, two Roscosmos cosmonauts will exit the @Space_Station to install a solar radiation experiment on the Zvezda service module. Learn how to watch live: go.nasa.gov/4wJqFoj

@elonmusk The hard part is building a culture that rewards truth instead of punishing the person who says it.

Critical feedback is much appreciated

@benln Worth watching, but the real value is hearing how each of them frames the bottleneck differently because that's where the next few years of product decisions get made.

Hands down, the best AI series right now CS 153 at Stanford, published on YouTube Speakers include Sam Altman, Jensen Huang, Satya Nadella, Andrej Karpathy, Ben Horowitz

@NASA The best space orgs understand that storytelling is part of the mission because public trust compounds when people can actually feel the work behind the science.

We won four Telly Awards! 🎉 These awards recognized live coverage of Artemis II, documentation of lunar geology training, and a video journey through exploration with our past, present, and future space telescopes: go.nasa.gov/4doMdPB

@GoogleLabs @PomelliByGoogle The interesting shift is that these tools stop feeling like demos when they help teams produce their actual artifacts, not just prettier prompts. That’s when the workflow starts to matter more than the model.

We believe our Labs experiments are at their best when they’re helping you create whatever you can imagine. We had some fun and put them to the test ourselves. Take a look at our Labs-inspired creations: 🧵1/5 A brand book that we created with @PomellibyGoogle.

@TheSixFiveMedia @Applied4Tech The interesting part now is whether enterprise AI leaders can move from infrastructure talk to workflow adoption that actually sticks, because that gap is still where most of the value gets lost.

Excited to share that @Applied4Tech CEO Gary Dickerson will join us at The Six Five Summit AI Unleashed 2026. Gary will join our analysts & other executives and industry leaders shaping the next phase of enterprise AI across infrastructure, cloud, edge, and workflows. August 25–27 Fully virtual Register: sixfivemedia.com/summit