Dec 25 will be remembered as the time when agents took over.
Also this: wtfhappened2025.com/
Hypernetworks are back. Technique to multiply context window size.
pub.sakana.ai/doc-to-lora/
OSS Context-as-a-Service will be huge once we settle on one.
Why is everyone building the same thing?
Vercel indexes its entire GTM memory on turbopuffer turbopuffer.com/customers/ve…
Another major ✅ Anthropic built C compiler with 16 CC agents consuming $20k of tokens.
What stood out:
1️⃣ Its so true that agents have no sense of time and will happily run useless tests for hours. Feels like something that can be trained into a model like they did with token efficiency.
2️⃣Lots of caveats - just like with the browser project at Cursor - its basically a useless version because of its performance. It took $20k to pass the test harness but it will take 1000x of that to march the 9s.
3️⃣ “We realized we need to break the work down”. Kind of struggle to understand how this keeps being rediscovered over and over again. I mean - how do you eat an elephant? With a spoon.
Nevertheless - its SO cool!
anthropic.com/engineering/bu…
This dude WROTE C compiler in 7 days 7y ago before all the Tab Tab Tab.
GitHub - rui314/chibicc: A small C compiler github.com/rui314/chibicc
Cockroach score: simplest way to make sense of SaaSpocalypse.
🪳score = Hours spent in UI / Avg revenue per user.
Low-ARPU/high-use software is undisruptable. It does not leave room for underserved users. Especially not with subsidized $$$ tokens.
Chart below ranks by this relative to 52w drawdown and flags FCF/SBC/buyback stuff.
Somehow i didnt know they had a continuous drone shot of Everest summit. Breathtaking. youtube.com/watch?v=0pIyIMqw…
🦀 Everything is crab 🦀
“Every program attempts to expand until it can read mail. Those programs which cannot so expand are replaced by ones which can.”
Codex = Cursor = Antigravity = Conductor
All AI coding apps have evolved into an inbox.
P.S. TIL that nature has evolved crabs multiple times: grokipedia.com/page/Carcinis…. Crabs dominate.
VCs taking medicine:
Funding 300% YoY deals at 100x ARR only really works if your fund is $15b and $10m seeds have negligible impact on IRR. What do you expect over long term? YoYs revert to the mean.
This leaves majority of VCs in an uncomfortable position between a rock and a hard place:
insights.euclid.vc/p/the-qui…
Had big arguments with my team about this since 2024. Most AI automation use cases in the real world do not need a database for one simple reason:
Automation - by definition - is focused on a task done manually by a person. People don't have a database in their brain.
So, don't add a database unless one is really needed for transactional processing.
llamaindex.ai/blog/files-are…
This is a major milestone. Cursor ran for 3 weeks uninterrupted to create a 3m line-repo with a simple browser using GPT5.2.
Browsers are notoriously difficult pieces of software to create (Chromium has 50m lines of code, Linux has 40m).
The fact that browser functionality is verifiable makes it a bit easier for coding agents, but it is a massive milestone nonetheless.
RL environments and verifiewrs is where it is at.
AgentFS: very interesting mountable filesystem on top of sqlite.
A lot of teams use databases where none are required. Using database is not free lunch - it inevitably raises complexity and complexity needs to be fought at every turn.
penberg.org/blog/disaggregat…
Reminder that SEAL snipers killed Somali pirates with 3 synchronized headshots by parachuting onto a Navy destroyer and then targeting a lifeboat in rough water thru its 1x1 ft ports at night.
This book from then-head of special ops is fantastic and full of sea stories. Under his command US was doing 10,000+ operations a year.
amazon.com/Sea-Stories-Life-…
Best reads on both sides of the AI argument after consuming copious amounts on the topic:
AI power leads to loss of control: gradual-disempowerment.ai/ and intelligence-curse.ai/
AI is normal tech and none of that happens: knightcolumbia.org/content/a…
I am on the "normalist" side.
This is what we automate for:
From 1965 to 1995, the typical adult gained six hours a week in leisure time (300 hours a year) and almost all of it went to watching more TV.
theatlantic.com/magazine/arc…
Treasure trove in a "2025 wrap" letter from DeepMind's Zhengdong Wang. Started clicking on links and spent 3hrs reading last night :)
zhengdongwang.com/2025/12/30…











