just spent hours coding with claude 3.7 sonnet, here is what I have to say

this is a serious violation of people’s basic right to privacy. they’re using these spywares to monitor us 24/7, tracking our every activity. this future is unacceptable. View quoted note →

Google secretly installed this app on various Android devices without users’ permission. Reports suggest that it can scan through your photo gallery and takes up 2GB of storage. You might want to check your phones; you can just uninstall it from the Play Store: https://play.google.com/store/apps/details?id=com.google.android.safetycore

Something intriguing is on the horizon! Stay tuned! 🕊️ View quoted note →

qr codes suck we need something better, something more secure, & better looking.

Now, let’s create a personality based on none other than Dostoyevsky. We can upload all his books directly into his personality, model's 2m context window can easily handle it. But you know what’s even more interesting? We can use an O3 model, enable deep research, and generate synthetic data specifically designed to recreate his personality. This wouldn’t just rely on his books but would also capture a deeper understanding of his literature and its impact & then use that synthetic data to train his personality. BTW, these personalities are shareable. So whenever I launch this app in the future, we’ll be able to share, import, and export them. In fact, the chat history and personality data stay on your device. You can even switch the API and use your own, it will still work. It’s more like a client.

One thing I feel truly fortunate about is that I have an opportunity to work on things I truly care about & want to see in the world. Half a decade ago, I heard someone say you should build the things you want to see in the world. I am currently in a position where I'm quite literally doing it. I get it, I'm 25, just getting started. But I really hope this remains the case in the future too.

This price for a multimodal model is insane. How do you even compete with this? These Gemini models are not only cheaper, they're 10x-20x cheaper than all other available options, extremely fast. With virtually no difference in answer quality, AI is becoming a commodity, faster than expected.

One interesting aspect of these LLMs is their general-purpose nature. eg, a model proficient in coding is likely also adept at writing or understanding data. Other factors, such as context window, also play a role. The model I'm using has a 2-million-token context window, roughly equivalent to 20-30 books per personality. Consequently, you don't need 10 different AI app to handle 10 different tasks. A single AI app, with its behavior modified, can handle all of them, often with a single model. View quoted note →

Today I built an app I had wanted to build for a very long time. It's basically a personality-based AI chatbot. You provide 2 simple things: 1st, instructions on how the model should behave. You can be very specific. 2nd, data. You want the personality to understand, that data can be bunch of PDFs. It then turns that data & instructions into a personality. So, next time you ask it anything, that personality will have all the context of the data you provided & will behave according to your instructions. When I'm all set, I'll open source it. you can simply attach your own API.

I want to build something like Blastr Proxy Relay. Any event posted on this will basically be broadcast to all major relays. It should also send a monthly invoice with a (pay-what-you-want) PWYW model instead of a fixed amount. When invoice is paid, the entire amount should be distributed among the relays it’s broadcasting to. Any thoughts?

This Twitch streamer has done something insane. He has managed to beat 7 FromSoftware games consecutively, all at level 1, without taking a single hit. 😳 He conquered are Demon’s Souls, Bloodborne, Dark Souls 1, 2, and 3, as well as Sekiro and Elden Ring, all within a single run. Those who are unaware, these are some of the most challenging games ever created.

Tested Grok 3, its “deepSearch” & “think” mode with premium+. First thought: they’re clearly manipulating the benchmarks. It’s not the best model by any stretch, especially in math & code. That much I can say for sure, so benchmarks in those areas don’t make much sense. XAI’s deep search is nowhere close to OpenAI’s deep research. In fact, I don’t think they’re even in the same category. XAI’s deep search is more like Perplexity’s but slightly better. That said, Grok 3 isn’t a bad model. It’s actually better than I expected—way better than Grok 2, but there’s nothing extraordinary about it. They also didn’t open-source it, so it doesn’t add anything meaningful to the ecosystem. Bonus: If you’re someone who wants to play with cutting-edge AI models & features, like models with chain-of-thought reasoning, grounded web search, Gemini 2 Pro, & the latest SOTA models before launch, with 2M-token context, all for free for personal, uou can use:

Google AI Studio

The fastest path from prompt to production with Gemini