That's how I read the situation too! The funny thing is that when ChatGPT released, he was using it all the time. I couldn't talk to him without him checking GPT before making any argument. In person.... Suddenly he stopped.
Started playing Eternal Darkness with a friend yesterday. A game released in 2002. We get stuck. I tell him to check ChatGPT. He checks YouTube instead. He’s not a technical person, he’s an airline pilot. What do you infer from that?
10/ And that they grow up looking up not with the dream of leaving Earth behind, but with the dream of continuing to explore our universe to better understand and protect the planet we all share.
Original photograph: www.nasa.gov/image-articl...
9/ My dad waited 54 years to see humans return to the moon. It took me 30 years to see them go for the first time. I hope new generations never have to wait that long again.
8/ Finally, in the lower right, Venus shines quietly. The zodiacal light seems to point directly toward it. A reminder that despite being our closest neighbor, it is anything but hospitable, like the vast majority of the universe.
7/ When we look toward the poles, both auroras glow timidly in the frame. A visible side effect of Earth's magnetic field at work. Our own force field, protecting us from solar radiation.
6/ On the right, partially veiled by clouds, we can make out Brazil. And despite our perception, we can also see how close it sits to Africa and Europe. Suddenly the Atlantic doesn't seem as vast as it usually does.
5/ The 1st thing we notice is the Sahara Desert imposing itself across Africa on the left. Below it, the Iberian Penins., with the coast of Portugal glowing bright. How many people were taking a sip of wine while being photographed from more than 250k km away without realizing it
4/ A shot like this was impossible the last time we had people up there. Apollo crews carried 160-speed color film. This was taken on a Nikon D5 at ISO 51200. A reminder of how far science can take us in a single lifetime.
3/ Look closely and you'll realize this is Earth's night side, lit not by the Sun but by moonlight. The Sun is entirely behind Earth, making this a kind of solar eclipse in reverse.
2/ Last week, we watched the first crewed mission head back to the moon in 54 years. If everything goes according to plan, two men may be standing on its surface in about two years.
The launch made me shed a tear. This photograph on their way there, made me shed another.
1/ Amidst so much change and uncertainty in the world, there's something grounding about waking up to a photograph like this.
My dad watched men land on the moon in 1969, when he was only 8 years old. He's the reason I care about space exploration.
This is my favorite climate change chart. Japanese monks, aristocrats, and emperors kept meticulous records of cherry blossom festivals for 1,200 years and accidentally built the world's longest climate dataset.
Prepping slides & demos for my Voxxed Days Amsterdam 2026 deep dive tomorrow
Packing in Memory, Advisors, RAG, Tool Calling & MCP, Agentic Patterns, A2A & ACP, LLM-as-a-Judge, Guardrails… and somehow fitting it all into 2 hours!
See you there 👉 amsterdam.voxxeddays.com/talk/?id=12905
The teams that win here won’t just build smarter agents. They’ll build better systems around them.
That’s the part a lot of people miss:
𝗔𝗜 𝗲𝗻𝗴𝗶𝗻𝗲𝗲𝗿𝗶𝗻𝗴 𝗶𝘀𝗻’𝘁 𝗼𝗻𝗹𝘆 𝗮𝗯𝗼𝘂𝘁 𝗴𝗲𝘁𝘁𝗶𝗻𝗴 𝗮𝗴𝗲𝗻𝘁𝘀 𝘁𝗼 𝗿𝗲𝗮𝘀𝗼𝗻.
It’s about making those agents behave efficiently inside real systems with cost, latency, and reliability constraints.
𝟮. 𝗜𝗻𝘃𝗮𝗹𝗶𝗱𝗮𝘁𝗶𝗼𝗻 A cache only helps if you know when the data is stale.
𝟯. 𝗖𝗼𝗼𝗿𝗱𝗶𝗻𝗮𝘁𝗶𝗼𝗻 If multiple agent steps need the same data, they should share it instead of re-fetching it independently.
The hard part isn’t calling tools. It’s managing the economics and behavior of tool use at scale.
That’s where 3 things start to matter fast:
𝟭. 𝗖𝗮𝗰𝗵𝗶𝗻𝗴
If the data was already fetched recently, don’t pay for it again.
You can give a stock analysis agent a tool that fetches market data from a 3rd party API, and in a demo that looks fine.
In production, if the provider charges per call and the agent keeps fetching the same ticker, you don’t have an agent problem.
You have a systems problem.
A lot of AI engineering content still assumes orchestration is the hard part:
- Which model?
- Which framework?
- How many tools?
- What memory layer?
But once agents touch real infrastructure, the bottlenecks start looking a lot more like classic software engineering.
Enterprise agents won’t scale because you gave them better prompts. 𝗧𝗵𝗲𝘆 𝘀𝗰𝗮𝗹𝗲 𝘄𝗵𝗲𝗻 𝘆𝗼𝘂 𝘁𝗿𝗲𝗮𝘁 𝘁𝗵𝗲𝗺 𝗹𝗶𝗸𝗲 𝘀𝘆𝘀𝘁𝗲𝗺𝘀.
Yesterday, @bsbodden.bsky.social and I delivered our Designing Multi Agent Systems with Spring AI hands-on lab at JavaOne in San Francisco. More than 35 people showed up and it was the lab with most registrations of the day. 🙏🙌☕️ #java
openai.com/index/introd...
The thing is that the longer the context window is the worse retrieval accuracy is as well.
For GPT 5.4 that was released last week, you can see that accuracy is ~97% up to 32k tokens, but drops to ~36% when over 512k tokens.
You can fit a million tokens, but GPT will not use it effectively.
That’s actually great even for extracting brand assets from the company you work for. Sometimes these assets are buried underneath layers of directories within Google Drive. This would be faster 😆
Isn’t this difficult due to the quadratic cost of the attention mechanism? Afaik attention decreases as context windows increase due to LLM engineers trying to trick the quadratic cost by implementing workarounds such as sliding windows. I haven’t paid close attention to recent researches, though
But the lesson is clear:
Ambition isn’t a strategy.
Shipping is.
Definitely a valuable lesson I’m gonna take with me for when I start my own startup in the future.
Next time, I’ll start smaller.
cc @sseraphini /7
And if this were a real startup competing in the open market, that would’ve been fatal.
Luckily, we’re on the same team and company. His solution is gonna serve me as much as mine would’ve served him. We can even build on top of his and possibly add all features I envisioned. /6
But his could ship immediately.
Mine was overengineered. It was harder to extend. Harder for non-technical people to contribute to. I spent hours trying to simplify it instead of having started simple.
I couldn’t ship an MVP. /5
@riferrei.com built one using GitHub devcontainers.
@guy.dev built something similar to mine, but much simpler.
His version had:
• An IDE
• A terminal
• A web page
• Instructions
That was enough.
Mine had more features. I tried to build a long-term vision in three weeks. /4
