AI 2027, Gemini 2.5 and ChatGPT Image Generation – Live and Learn #65

✨ Quote ✨

If you insist that anything too common, anything come by too cheaply, must be boring, then all the wonders of the Singularity cannot save you. You will grow weary of green wine and sick of crimson seas. But if you can bring yourself to really pay attention, to see old things for the first time, then you can combine the limitless variety of modernity with the awe of a peasant seeing an ultramarine mural.

– Astral Codex Ten - (source)

This quote and the whole accompanying blog post hit hard. It's how I try to live: to appreciate the mundane, but not so-mundane things. Everything, if looked at closely enough is its own marvelous thing. Even something as simple as the color blue and that we have copious amounts of it, would be magical to people living in Medieval times. We often forget that we live in a world of beautiful marvels. But the only thing we have to do is to pay close attention once again, and the world transforms into one where wonders abound at every corner.

🖇️ Links 🖇️

AI-2027 by AI Futures Project. This future scenario took a while to read but I highly recommend it. It's the best guess of some brilliant people of what the future is going to look like when AI development continues to be more and more automated. The outlook they paint reads like crazy science fiction, but don't let that fool you. This is serious work of some serious forecasters, trying to predict what reality is going to be like a few years from now. After 2027 they agree that things become speculative and harder to forecast because of the sheer speed that AI development will reach by then. They even split their scenario into two different potential outcomes: a race to the bottom and a slower, more controlled approach. Reading this I have to remind myself, over and over again that this is not science fiction but "just" an extrapolation of trendlines. My friend has this question on his wall: What would you do if this were your last year on Earth? It used to be a rhetorical question to get you thinking about how to best spend your time, but soon it might become a very real question. And we will have to find our own answers as to how to spend our time in the prospect of imminent ASI and a level of progress so steep that it boggles the human mind. For more info about this AI 2027 scenario, you can also listen to the Dwarkesh Patel Podcast, where he interviews the people who wrote this scenario.

Gemini 2.5 by Google. Gemini 2.5 is Google's newest flagship model. It has impressive benchmark scores and Twitter is going wild with using it to vibe code and do all the LLM stuff. To me, the biggest differentiator is that Gemini can use real-time video input as well as YouTube videos for contexts.

FOSS Infrastracture under Attack by The Libre News. This article dives into how AI bots by big companies and startups from the US and China alike are crawling the entire internet. Over and over again, with zero regard for robots.txt disallow rules. At this point, it is fair to say that AI bots are DDOS attacking public infrastructure. Server owners and administrators have to bear huge costs for serving those requests. Some of the sites see 80% of their traffic come from AI bots and have to pay thousands of dollars to serve those requests. In response, people started to adapt tools like Iocaine or Anubis to send the crawlers on wild goose chases across endless link mazes or make them solve hard computational problems that incur a heavy cost on every request sent. There's another good post detailing more of what is going on. Also thanks to Jonathan for sending me this!

ChatGPT Image Generation by Open AI. ChatGPT now has a new image generation model embedded directly into it. You can chat with GPT-4.0 to create, edit, and manipulate images at will, all through the familiar chat interface. It's really good at handling text and the image quality is state of the art. But it still breaks down when trying more complicated things (like rendering a whole periodic table of elements with cartoonish drawings). It also hallucinates, like all LLMs. Open AI published the model system card over here diving into some of the security aspects of a powerful image generator like this. So far, the main thing the internet has used this new ChatGPT feature for is to "ghiblify" their photos, spawning a viral trend, resulting in millions of new signups to the OpenAI platform.

Brain to Voice Neuroprosthesis by UC Berkeley. Neural implants are getting better and better. This one can decode what the patient wants to say and produce sounds mimicking their original voice. This works in real-time, with a chip implant reading their brain activity! The whole idea is still in its infancy and the generated voice feels robotic and falls into uncanny valley territory... But still! That something like this is possible at all is basically magic.

🌌 Travel 🌌

I'm in Bolivia now, exploring the country for a month, starting with La Paz. The altitude here makes it hard to breathe and I struggled with Acute Mountain Sickness (AMS) after I arrived. The people here call it Sorojche. Luckily the pharmacies sell Sorojchipil, essentially Acetazolamide mixed with Ibuprofen. It helps to fight the headaches and makes the body adapt to the lower oxygen concentration by breathing more rapidly. I am already feeling a lot better now, but still have to take things slowly.

My favorite part of La Paz, by far, is their cable car public transport system called Mi Teleferico. It's cheap, silent, and you don't have to wait. Plus, you get to see the whole city from above ^^ But see for yourself in the photos below.

Over the next few days, I'll explore the Salar de Uyuni and mountain salt lakes in the South towards the Border of Chilé 😊✌🏻 See you again in two weeks and until then, take care!

🎶 Song 🎶

Laughing With by Regina Spektor

Youtube Music | Spotify

---

That's all for this time. I hope you found this newsletter useful, beautiful, or even both!

Have ideas for improving it? As always please let me know.

Cheers,

– Rico