1.
What a good agent eval looks like
A benchmark should measure recovery from stale evidence and conflicting instructions, not just raw completion
Backlist — 28 Jun 2026 UTC
Strongest crop skewed AI/infra, but the slate was kept mixed with archaeology, architecture, economics, design, hardware, and biology.
6.
Why intelligence should not be artificially expensive
Cheap open models are turning intelligence into a commodity input instead of a luxury feature
11.
Low-rank directions across checkpoints
Finding a stable low-rank direction across checkpoints hints that model merging is only one piece of a bigger geometry
12.
LLM software is AOT for now
Prompted code generation is starting to look like ahead-of-time compilation, not interactive software
14.
PCB paste geometry as a yield lever
Paste geometry at the PCB level can change yield and assembly quality in ways most software people never see
15.
The first fully virtual Herculaneum scroll
Virtual unwrapping finally read an entire carbonized Herculaneum scroll without opening it
24.
A student formalizes an Erdős problem in Lean
A student formalized an Erdős problem in Lean with an unusually large proof, which is a big deal for machine-checked math
27.
A Suminagashi simulator
Simulating Suminagashi turns a traditional ink technique into an interactive generative-art tool
28.
Figma can spin up shader plugins on the fly
Figma now lets users spin up motion-aware plugins on demand, blurring design and extensibility
30.
Can neural nets be written as one equation?
A global equation for neural nets would change how people think about learning dynamics and AGI
32.
2nd gpumode competition for the linear algebra series is out symmetric eigendecomposition (tridiagonalization -> …
2nd gpumode competition for the linear algebra series is out symmetric eigendecomposition (tridiagonalization -> tridiag eigensolve -> back transform), qr_v2 is also on its last stretch final <48 hours keep the submissions coming
36.
Effective altruism is an amazing nerd trap because they are right about most things but have really harmful socia…
Effective altruism is an amazing nerd trap because they are right about most things but have really harmful social dynamics. 100% custom built for a particular neurotype
37.
this means ur agent wall clock is about to be dominated by tool calls like find btw
this means ur agent wall clock is about to be dominated by tool calls like find btw depending on price this raises some very interesting questions around auto verification loops/etc
43.
the token-maxxing era is quietly ending
the token-maxxing era is quietly ending every company i talk to now wants to use fewer tokens, not more.
46.
San Francisco is the university network for the drop out entrepreneur
San Francisco is the university network for the drop out entrepreneur
51.
If people doing frontier science in Unis have no longer access to frontier models and those in industry do, does …
If people doing frontier science in Unis have no longer access to frontier models and those in industry do, does university research remain viable?
54.
The Professor's Laboratory
The Professor's Laboratory
61.
I'm not even sure we *should* decarbonize things like steel/cement… might be better to just develop low-cost CDR …
I'm not even sure we *should* decarbonize things like steel/cement… might be better to just develop low-cost CDR and massively scale it up
69.
google should be the tech giant building a prediction market, not meta
google should be the tech giant building a prediction market, not meta do you know how easy it is to monetize someone who googles a future-looking result and disagrees with it? google could: - spin up non-obvious but in-demand markets ba
76.
Experimenting with xr to control the b601 with an overlay and servo position
Experimenting with xr to control the b601 with an overlay and servo position
81.
if anyone wondering, they asked the value of 5C2, term independent of x ( typical jee binomial qsn), and area of …
if anyone wondering, they asked the value of 5C2, term independent of x ( typical jee binomial qsn), and area of ||gm given line eqns, a poisson dist. qsn and some prob stuff, and one determinant with adjoint qsn. ML consisted of pca, ove
83.
Uzun zamandır newsletter yazıyorum ve platformlarını kullanıyorum.
Uzun zamandır newsletter yazıyorum ve platformlarını kullanıyorum. Son zamanlarda Ghost + Claude Code kullanıyorum bülten altyapısını organize etmek için. API desteği ile neredeyse her şeyi yaptırabiliyorsunuz. Ve sadece text mesajla. Yen
86.
Yes, GPT-5.6 is still Spud, trained further
Yes, GPT-5.6 is still Spud, trained further
90.
Why? “Mobile” = little screen and no keyboard, which is just strictly worse. A human thing, not a software thing.
Why? “Mobile” = little screen and no keyboard, which is just strictly worse. A human thing, not a software thing.
How it ranks: Backlist reads my Twitter/X timeline, scores every tweet for substance with an LLM rubric (not engagement), and publishes the daily top picks with a one-line takeaway. Curated by Surya Dantuluri.