Dancing on the event horizon
Don't trip.
🤖 AI
On 60 minutes, Demis Hassabas said that we are 5 to 10 years away from AGI. He claims that by 2030 we will have "a system that really understand everything around you in very nuanced and deep ways and is kind of embedded in your everyday life". If this, or some variant of the slowdown branch of AI 2027, pans out the world we live in will be very different very soon. Sometimes I wonder if it makes sense to work on anything that isn't pure art. Its clear to me that we are near the event horizon because every day, its getting harder to see what tomorrow looks like. I'm looking for more ways to think about this, shoot me a note and let’s chat about it.
Thinking with images: OpenAI introduced o3 and o4-mini visual reasoning models that are capable of thinking with images in their chain-of-thought. When you add MCP/tool use to reasoning multi modal models some very interesting things can start to happen since the models can start to pull in data it needs --or even interact with the outside world. I think that reasoning + tool use is the main driver behind the success of Cursor / Windsurf. Bringing this to other jobs outside of code could be pretty powerful. I've been experimenting with giving reasoning models access to internal and external systems and see a lot of potential. I hope to be able to share some findings soon.
Windsurf may get bought by OpenAI for $3b. Not sure what to say on this without sounding like a hater. Even though they have a lot of revenue and users, I'm just not sure how sticky these customers are. I talk to developers every few days who are switching between Cursor, Windsurf, and VSCode's Insider version consistently. The only reason I can think OpenAI is looking to do this is to defend against Anthropic (by way of Cursor) dominating the coding game. Even then, $3b? I guess if the deal is in stock so its really all just house money.
Gemma 3 QAT brings state-of-the-Art AI to consumer GPUs: Many of the powerful open models (DeepSeek R1, Llama, etc) require multiple $25k+ GPUs to run. This just simply isn't accessible to anyone outside of the hyperscalers. Typically you would sacrifice performance to get LLMs to work on less powerful hardware but Gemma 3 27B is scoring well on benchmarks. This model can be run on consumer-grade GPUs like the RTX 3090 which cost $1700. The 3090 is still expensive, but we are trending in the right direction.
🔗 Cool *hit
A unique sound alleviates motion sickness: Researchers at Nagoya University have developed a sound stimulation device that significantly reduces motion sickness by stimulating the inner ear with a specific 100 Hz sound wave.
PiLiDAR/PiLiDAR: The PiLiDAR project is a DIY 360° 3D panorama scanner using LiDAR technology. The hardware setup is based on a Raspberry Pi 4 and includes a HQ camera and stepper motor for movement.
A school in Austin called Alpha School is teaching its students in a 2 hour learning model using AI with human facilitators. During the rest of the day the kids work on developing life skills. The kids are even running a food truck. Too bad its in Austin.
🔈What I'm listing to
James Currier's ideas on network effects are great but the most interesting takeaway for me is this quote on technology windows:
"Between 94 and 2013 it was a fantastic time to start companies in the consumer software space and invest in from seed or series A. In 2014 the window just closed You can look at the number of unicorns created in two from 2014 to today And it's it's really small It's like Discord, TikTok, Starlink, ChatGPT there's like 10 of them in the west versus how many you know during the the previous let's say like 10 years you might 10x that number...15 a year"
I've been saying something like this for years. Tech today just isn't the same as it was in 2010 but founders, investors, etc are all thinking about companies as if they were built in 2010 --two Sanford cofounders one technical/one business, 0 to 100m DAUs in the first year, low churn as PMF indicator, etc. When you look at the biggest latest successes today, like OpenAI or Starlink, the structure / G2M doesn't line up at all with what you saw back in 2010. OpenAI was a non profit research org where not all of the founders were involved in the day to day operations. Starlink and TikTok are subsidiaries of other companies. This has me wondering what are we getting wrong in consumer tech these days —and what opportunities exist because we are getting things wrong.
📚 What I'm reading
Do Androids Dream of Electric Sheep?: For those that don’t know, this is the book that became the movie Blade Runner. I've had an on again and off again relationship with this book for years. But "this time is different". I've gotta finish it before AGI and it sounds like I only have a few years left. No time like the present.