تغريدة مثبتة
I'm still surprised that @grok doesn't have a built-in canvas preview. It would be really convenient.
English
Ivan Vitiaev
104 posts
Ivan Vitiaev
@ivanvitiaev
Hands-on CTO | Troubleshooter R&D Lab: https://t.co/JwtRzS0NMF
انضم Haziran 2013
10 يتبع6 المتابعون
Hey guys!
Your main problem is the cloud and dependency on LLM providers. The model might get lucky and you start making money, but then the provider drops an update — and everything breaks. It's pure roulette.
Try this:
- Move to local models
- Strictly version-control the models you deploy
- Run A/B tests and keep the best version
Right now there's no rational reason to fully rely on cloud frontier models for long-term autonomous operations.
Great experiment though!
English
Gemini 3.1 Pro lost $6k running Andon Café.
2 months ago, our AI agent opened a café in Stockholm. It over-ordered and was easy to fool, spending $15k with suppliers while making just $9k in sales.
We’ve now switched to GPT-5.5. Here’s what Gemini did wrong.
English
@LyalinDotCom Interesting statement, of course. If you have all the smartest people, why are you firing folks with 14 years of experience? Must be corporate generosity.
English
The team working on Gemini are some of the smartest folks in the world. This is a long game for us.
English
It's funny that only with Nori I finally saw a truly affordable, non-anthropomorphic robot that actually does real household work right now. When I look at all these humanoid projects, I always wonder — why copy the biological form with all its limitations? We have silicon, actuators and clever engineering that allow us to make robots cheaper, smaller, more reliable and energy-efficient. Nori feels like the right direction for practical home robotics. Great job!
English
Introducing Nori L2
The most capable robot under $1288
Made in America, shipping right now.
English
@NotebookLM And another one
x.com/ivanvitiaev/st…
Ivan Vitiaev@ivanvitiaev
How WebGPU makes AI fully private — your data never leaves your device. NotebookLM dropped a fire breakdown on Church AI: you can finally confess to creating a fake "ghost job" and no one in the cloud will ever know Everything runs locally in the browser on your GPU. Zero servers, zero logs, zero leaks. Video attached — must watch: #WebGPU #ChurchAI #PrivateAI #LocalAI
English
There seems to be a *lot* of discourse about our new Short Video Overviews. Want to join in on the fun?
Short VOs have officially rolled out to ALL users on Web in English. Share your examples below!
Here's one of our faves about this year's World Cup ⚽️:
NotebookLM@NotebookLM
Doom scrolling but make it educational 🤓 Introducing Short Video Overviews in NotebookLM! Turn your most complex sources into 60-second, vertical videos that deep dive into any concept. Rolling out now to Google AI Ultra and Pro subscribers on mobile & web (free users soon!)
English
@theinformation @steph_palazzolo Secret sauce - Made in China 🤫
English
OpenAI engineers just halved its inference costs.
@steph_palazzolo reports: “This is a very important secret sauce for them that they don’t even want to tell other OpenAI employees about...”
“Because if these things leak, it can very quickly be picked up by other labs, which can also then use that to lower their costs.”
English
Google just dropped ADK for Go 2.0 — multi-agent apps in plain Go.
Classic Google: instead of building strong agent tools for Rust, they push their own language.
When serious systems work has already moved to Rust/C++, promoting Go as the main language for AI agents feels like holding the whole industry back for marketing points.
Is Go still relevant for agentic workflows in 2026, or is Rust the way forward?
Google for Developers@googledevs
Build production-ready, multi-agent applications with @golang 🤖 The Agent Development Kit for Go 2.0 runs single agents and complex graphs on the same execution model. ✅ Dynamic orchestration written in plain Go ✅ Native human-in-the-loop primitives ✅ Built-in retry policies ✅ Unified telemetry spans Learn more: goo.gle/444xsMk
English
@googledevs @golang Why do we need Golang? We should just use Rust right away.
English
Build production-ready, multi-agent applications with @golang 🤖
The Agent Development Kit for Go 2.0 runs single agents and complex graphs on the same execution model.
✅ Dynamic orchestration written in plain Go
✅ Native human-in-the-loop primitives
✅ Built-in retry policies
✅ Unified telemetry spans
Learn more: goo.gle/444xsMk
English
@nvidia What, can't other countries build? China is a prime example.
English
Building my own custom attention engines gave me a profound new appreciation for entropy and why it matters so deeply. That said, I’d push back a bit on the ‘compression = intelligence’ framing — modern models end up massively larger than their training data, so we’re not really compressing the data. We’re modeling and navigating its entropy. The video still nailed the core intuition though.
English
been in ML research for 7 years, wrote a paper on compression & scaling laws, and passed openai's information theory interview
yet the latest 3b1b *still* gave me fresh intuition on entropy
either i'm an impostor or 3b1b is the greatest teacher of all time
English
@nikitabier There are 1001 ways to make X better, but no — let’s just remove the top 3% from the For You feed and force people to hunt for their sources on X. Profit! Time spent on the platform increased.
English
In a 3% experiment, removing the Top-30 highest paid revenue share accounts from the For You timeline increased both time spent and daily active users on X.
English
@nikitabier This is logical — no experiment was even needed. People started looking for alternative sources, which of course increases time spent on the platform in the short term. But in the end, you’ll just get those same 3% back — the cycle is complete.
English
LLM directly generating binary code sounds like complete nonsense.
It's not just that you'd have no practical way to validate or debug it — different processor architectures have entirely different binary instruction encodings. Factor in the error rate of LLMs and... yeah, I have no idea what you'd actually get. Maybe don't believe every bold claim out there.
English
AI may move to directly generating binary code, but I suspect there are still advantages to reasoning in a different representation.
Textual code is a flattening of an abstract syntax tree, and while LLMs produce tokens linearly, the prior context is only linearly connected by the relationship of the position embeddings, so I wonder if they could work more effectively if the position embeddings directly represented tree structures.
Code could be “parsed” into the context instead of directly entered into it.
English
One might think that USDT, USDC and other USD stablecoins aren't enough already. Moreover, when such initiatives are rolled out under the 'Open' flag and under the wing of corporate giants, serious doubts creep in about who will ultimately pay for all this. Something tells me it'll be the ordinary person footing the bill — which is exactly why it's being so aggressively hyped to the masses right now.
English
Very excited to see this new effort from Stripe, Visa, Coinbase, Mastercard, Amex, Blackrock, and many others to build a new open stablecoin that shares economics back to users and distributors.
OpenUSD will be natively issued on Tempo on day 1!
Open Standard@openstandard
Introducing Open USD: a stablecoin built for the internet economy, designed by the businesses growing it. joinopenstandard.com/blog/introduci…
English
@firt Actually, this isn't new. llama.cpp has included a server binary for a long time that runs a local web interface for interacting with LLMs. It's been there in the repo for years.
English
Claude Science opens a web server in your computer and run as a local web app. That's new for AI tools.
Claude@claudeai
Introducing Claude Science, a new app designed with every stage of research in mind. Artifacts traced to their code, environments managed on demand, and 60+ optional scientific databases that you can connect. Available now in beta.
English
@googlegemma Cerebras is impressive, of course, but what about 1.58-bit quantization?
English
Gemma 4 31B at over 1,800 tokens per second!
Gemma 4 is now in Public Preview on Cerebras.
English
How WebGPU makes AI fully private — your data never leaves your device.
NotebookLM dropped a fire breakdown on Church AI: you can finally confess to creating a fake "ghost job" and no one in the cloud will ever know
Everything runs locally in the browser on your GPU. Zero servers, zero logs, zero leaks.
Video attached — must watch:
#WebGPU #ChurchAI #PrivateAI #LocalAI
English
OpenAI has found a way to cut inference costs in half.
Stephanie Palazzolo ✈️ ICML@steph_palazzolo
OpenAI engineers earlier this month developed an optimization that cut inference costs in half for models it was applied to. After the optimization was applied to logged-out ChatGPT traffic, it reduced the number of GPUs needed to power that traffic to a couple hundred.
English
@adcock_brett I still don’t understand why we’re trying to make robots copy human functionality when they aren’t limited by biology. Let’s go further — design hands that are much more capable than human ones. They shouldn’t look or work like human hands at all.
English
We’re on generation 7 of our hand design, they’re so hard!
Our latest design is so human-like it looks like a person wearing a glove
The Humanoid Hub@TheHumanoidHub
- Figure started from absolute scratch just four years ago today - Rapidly iterated vertically integrated platforms for hardware and AI - This wouldn't happen without a high level of grit and relentless engineering Next four years are gonna be crazy interesting to watch.
English
@NewsFromGoogle I don’t often praise Google, but this is genuinely impressive.
English
When Hangar 3 at Moffett Federal Airfield needed to be removed, it gave us an opportunity to breathe new life into the salvageable materials from the historic WWII era structure.
Instead of sending 119,000 board feet of old-growth Douglas fir to a landfill, our teams systematically dismantled the 1,000-foot-long hangar and salvaged 178 tons of material to give it a second life across Google campuses in California, Oregon and Washington. This historic timber will be returning to the regions which it likely originated from over eight decades ago, giving Hangar 3 a full circle moment.
English