philipplgh (@philipplgh) header
philipplgh (@philipplgh) profile picture

looking at things

Joined October 2013
Photo by philipplgh (@philipplgh)Photo by philipplgh (@philipplgh)Photo by philipplgh (@philipplgh)Photo by philipplgh (@philipplgh)Photo by philipplgh (@philipplgh)Photo by philipplgh (@philipplgh)Photo by philipplgh (@philipplgh)Photo by philipplgh (@philipplgh)Photo by philipplgh (@philipplgh)Photo by philipplgh (@philipplgh)
Can this be used to predict where the frontier is going? Partly. Backtested 3 times: the drift kept its direction every time, and 3 of the top 10 predicted risers were right (chance: 0.7). How much: worse than guessing no change. Where it points next:
You can ask @claudeai to map scientific papers based on SPECTER v2 embeddings. I mapped 29k arXiv papers on AI agents (2025–2026) by their embeddings to see where the frontier is moving. Agent research is turning from "can it?" to "why did it break?" 🧵
2
5
256
I wondered if we could tree-shake Chromium out of Electron apps. Most apps only use a fraction of Chromium's features. Could we strip out the rest? Slimium explored this but pre AI. Maybe coding agents could go further with fuzzy, non-deterministic dependency analysis. I tried compile-time flags first (see image). Result: just 3.6% smaller. Maybe not worth the effort.
173
AI coding is making people rediscover software engineering: diagrams, design reviews, planning, formal methods. But the most overlooked discipline is requirements engineering. It's not just asking more questions. You can't verify correctness against contradictory requirements.
177
The biggest waste of AI tokens might be writing the same code over and over again. Write it once. Verify it. Reuse it forever. Everything else becomes composition. Blockchain coordinates what gets built. Save the compute for gaming and running the Matrix.
5
231
As the AI divide widens and the world turns into the Hunger Games, everyone is looking to start a physical business catering to the wealthy elite. But before you open that luxury delicatessen, remember: even billionaires can’t buy a second stomach.
192
Funny how data hoarders are becoming AI overlords. Old bookmarks, passwords, devices, journals, personal notes, source code—all of it is suddenly useful again thanks to AI. While we're at it: don't forget to back up your sessions for future session mining. Or at least extend Claude's auto-delete timer.
1
3
246
Pro tip: backup scratchpads as well
2
61
I imagine anyone capable of profitably commanding 10,000 agents simultaneously has already left the clearnet. Literally zero incentive to brag, debate, or engage with an increasingly "permanent underclass" and regulators.
7
409
I see generative art onchain as an immutable testament to progress over time. The fact that what once took years of learning, sweat, struggle, and creativity is now trivial to produce paradoxically makes those early works even more valuable. Just like art throughout history.
1
1
7
374
This would be the ultimate capitalism-maxxing move now btw. Microsoft, hire me for more ideas.
1
6
628
Zooming into the tip of the arrow: 20 subtopics. The core coding-agent topics grow slowest. What grows is the layer around them: effect contracts, memory that revokes stale facts, proofs, audits of the evals themselves. 7 of the 20 had under 5 papers a month before 2026.
1
67
The full list, plus what lies between the subtopics. I expected empty space around the frontier. There is none: even the thinnest seam is nearly as dense as its neighbours, and in a backtest, thin seams didn't fill faster than full ones. The frontier grows as one block.
43
You can ask @claudeai to map scientific papers based on SPECTER v2 embeddings. I mapped 29k arXiv papers on AI agents (2025–2026) by their embeddings to see where the frontier is moving. Agent research is turning from "can it?" to "why did it break?" 🧵
4
1
5
597
How: arXiv search for "agent" in cs.AI, Semantic Scholar's SPECTER v2 embedding per paper, k-means with 40 clusters, share per month. Topic names are mine. The average paper drifts steadily away from Jan 2025, far above noise. No single jump.
1
67
Caveats: volume isn't quality, 2026 papers have no citations yet, and "agent" meant RL agents before it meant LLM agents. What should I weight it by next: citations, or which labs got there first?
49
A huge amount of what we call engineering progress is actually accumulated historical contingency, institutional inertia, vendor incentives, arbitrary conventions, and human cognitive limitations masquerading as technical necessity. We've confused three things: 1. Fundamental complexity — constraints imposed by physics, mathematics, and the actual problem. 2. Accidental complexity — constraints imposed by historical decisions, incompatible standards, legacy systems, and organizational habits. 3. Manufactured complexity — constraints deliberately preserved because they create competitive moats, professional gatekeeping, switching costs, or recurring revenue. AI could become extraordinarily effective at separating these. And once you can reliably distinguish them, much of modern engineering starts looking absurd.
3
3
272
5. Electrical engineering: legacy interfaces everywhere USB variants, power connectors, voltage conventions, industrial communication protocols, proprietary charging interfaces, incompatible smart-home ecosystems. Some diversity reflects real physical needs. Much reflects independent evolution and commercial incentives.
1
43
6. Building construction: the civilization-sized inefficiency Plumbing, HVAC, electrical wiring, insulation, structural systems, and building controls are often designed and installed as separate disciplines. The result is a vast collection of coordination problems, incompatible components, site-specific improvisations, and labor-intensive processes. Imagine optimizing buildings as integrated physical systems rather than collections of professional trades and vendor catalogs. There's a common mechanism here: human civilization has been optimizing locally with limited intelligence and limited coordination for centuries. The result isn't a globally optimized technological civilization. It's an enormous archaeological site of decisions.
40