Petros

@nikitabier @BillyM2k Aren’t you worried about generating negative social proof?

@BillyM2k The reason I’m interested in doing this because people overestimate how many of the followers are still active, especially if they grew in a prior era (e.g., COVID, etc)

Would it valuable to know how many of your followers have been active on X in the last 24 hours?

@sashakryzh @Baconbrix It works fine except when the update takes like 60s to download

@____petros @Baconbrix what about this flow: > app opened > check if update available > user presses "Get Started" (the first button on onboarding) > loading state on a button + expo-updates downloads and reloads the app > automatic navigation to the second screen

What’s the point of instantly generating software if you have to wait days for your users to get the update… Expo OTA makes mobile iteration speed as fast as the web. It’s pure advantage, easy to setup, and used by thousands of top apps.

@Baconbrix A new user doesn’t have time to benefit from this, they download the app and open it and boom they’re on a really old version of the code. We could make them all wait for the update but on some devices it can take a long time to download… any other solutions?

@____petros Setup background updates #improved-background-tasks" target="_blank" rel="nofollow noopener">expo.dev/changelog/sdk-…

@jmeistrich Can it accurately scroll to end where “end” is the TOP of the last item in the list? (Not the bottom)

Legend List 2.0 is LIVE! The ⚡️ fastest React Native list library, in 100% JS, now with more accuracy and speed! ✨Perfectly accurate scrollToIndex ✨Even faster ✨Zero mount time on new arch ✨Sticky headers ✨estimatedItemSize is optional ✨New and old arch

@zhiyuan_zhou_ @seohong_park @qiyang_li @svlevine Sounds like this but you haven’t cited it? arxiv.org/pdf/1910.02876

@seohong_park For more, see Paper: arxiv.org/abs/2507.07969 Website: colinqiyangli.github.io/qc/ Code: github.com/colinqiyangli/… This is led by the amazing @qiyang_li and with @svlevine

Action chunking works really well in imitation learning, and is essential to learning good BC policies in robotics. Can/should we apply the same idea in RL? We find that RL in the action chunk space, when done right (we call it ✨Q-chunking ✨), can be highly efficient🧵👇

@lp_rhodes @alexalbert__ How you mean, how do you use MCP to mirror functionality of a slack integration?

@____petros @alexalbert__ Salesforce says no 😞 MCP servers with rate limits seem to be the main solution for now

We’re rolling out more updates to Claude Code. First, the GitHub Actions integration is now available to Pro and Max plan users (in addition to pay-as-you-go API users).

@OfficialLoganK @geminicli Please let us call it by doing @gemini in GitHub?

Say hello to the @geminicli, a local CLI to help you build and maintain software with 1,000 free Gemini 2.5 Pro requests per day : )

@_avichawla Why do you pick the best of 2 responses instead of getting an LLM to mark a single response and only save if it gets a high enough mark? Picking the best of 2 will mean you sometimes still let bad data through? Is there an advantage

A detailed explainer on generating your own LLM fine-tuning dataset👇

@tannerlinsley @tan_stack Does this type of prefetching work on safari or still doesn’t work?

Coming soon: @Tan_Stack Router's `intent` preloading can now predict where your users cursor will go (you can even customize proximity) to preload the routes that users are most likely to visit *before* they even hover the link. Prefetch smarter, not harder.

@rauchg @thdxr Prefetching doesn’t work on safari right? Or does it now?

@thdxr Have you tried next-ai-news.vercel.app? Everything is dynamic, zero caching, fetches from Postgres, all RSC, auth / identity on every page. If a navigation is slow it’s one or more of; - They’re not prefetching - They don’t have fallbacks / loaders - They have slow backends

it's getting to the point where i can just feel that an app is using RSCs when i use it - nearly 1s page navigations i'm sure it can be improved but yet to come across an RSC app in the wild that feels good about to say v0 but just checked idk if they're using them anymore

@emilwidlund @letstri I assume this doesn’t work on safari though?

Today I learned that prefetch for <Links /> is defaulted to false when using the App-router. But true when using the Page-router. Thanks @letstri for shining light on the performance issue when navigating. Feels a lot snappier now.

@deliprao Yes they need to stop this @OfficialLoganK

Gemini Pro is such a good model if only it did not fail on every other query with "<FinishReason.RECITATION: 'RECITATION'>". Google folks you are killing the product with safegaurds.

@tedx_ai Which MCPs you using

Cursor + Sonnet 3.7 + MCP feels like being in a self-driving car for the first time 😇 All that's missing is more speed + parallel agents ⚡️

@OfficialLoganK @kimmonismus Being able to fine tune Gemini flash 2

@kimmonismus what would make you excited about Gemini? we have successfully shown over the last year we can keep pushing frontier performance, while maintaining the lowest cost, and adding new capabilities

1. Overall, Gemini 2.0 is kinda disappointing 2. pre-training hit a wall. 3. inference scaling is the new law

@kwindla @Shawnryan96 @pipecat_ai Does the api know about and can suggest YouTube video links? I know the app does but not the api

You can access all the publicly released model capabilities via the API. That includes images and video, and it can access URLs via the built-in search grounding tool. To build apps that tie everything together -- text, voice, image, video -- you can use the @pipecat_ai realtime agent framework.

Google's full release of Gemini 2.0 Flash is a great thing for the voice AI ecosystem. Up to this point, almost every production voice AI agent has used GPT-4o. Voice AI apps need an LLM with fast TTFT, good instruction following, reliable function calling, and natural conversation dynamics. GPT-4o was ahead on all of these attributes by a wide enough margin that using any other LLM mostly didn't make sense. Gemini 2.0 Flash is competitive with GPT-4o on all four of those attributes. Now we have two excellent models for voice AI. Competition is good!

@HelovesF1 @SecrtAgntSquirl @_akhaliq Thank you but I can’t seem to see where to create an account on their website? Have you found it?

@____petros @SecrtAgntSquirl @_akhaliq Heading over there too 👍👌🐇

ByteDance announces Doubao-1.5-pro - Includes a "Deep Thinking" mode, surpassing O1-preview and O1 models on the AIME benchmark. - Outperforms deepseek-v3, gpt4o, and llama3.1-405B on popular benchmarks. - Built on a MoE architecture, with activated parameters far fewer than those in the above models. - Achieves a 7x MoE performance leverage—delivering dense model performance with just 1/7 of the activated parameters (e.g., 20B activated params = 140B dense performance). - Engineering-wise, features heterogeneous system design for prefill-decode and attn-fffn, maximizing throughput under low-latency requirements.

@HelovesF1 @SecrtAgntSquirl @_akhaliq Where can we use it?

@SecrtAgntSquirl @_akhaliq It’s posted above too but here you go 🙏👍👌 Also, pricing-wise it is pretty incredible too: $0.022 per million cached input tokens, $0.11 per million input tokens, and $0.275 per million output tokens.

@leeerob How do you deal with speculation rules not being available on safari?

btw this is how vercel​.com works. We have 10+ microfrontends that stitch together into what feels like a single monolith. e.g. marketing site, microsite, docs, app + with the Speculation Rules API, you can still keep page transitions fast.