1.
A Lean-Checked Ethereum Verification Benchmark
Lean-checked tasks turn Ethereum verification into a benchmark you can actually trust
Backlist — 24 Jun 2026 UTC
Balanced technical research, market structure, policy, hardware, and a few durable startup/product stories; skipped most congratulatory replies and near-duplicate AI launch chatter.
2.
Why Direct Commodity Trading Matters for Futures Markets
Direct access to compute, metals, and other underlyings would make futures markets more useful for real hedging
3.
ARTS: Search and Test-Time Training for Huge Idea Spaces
The system combines exploration with test-time training in a way that could generalize beyond one paper
4.
OpenThoughts-Agent Opens Up the RL Recipe for Coding Agents
The paper lays out a data recipe for agentic coding that is detailed enough to evaluate and reuse
5.
Self-Play Plus a Pinch of Human Data Makes Driving More Human-Like
Thirty minutes of human demonstrations regularize self-play enough to produce much more human-like driving policies
6.
Transformers Still Hit Plasticity Loss
The paper says long training eventually erodes a model’s ability to adapt, even in stationary pretraining setups
7.
Qwen Builds a World Model for Seven Agent Domains (x.com)
Environment simulation becomes the training objective instead of a post-hoc hack for agents
8.
A Benchmark for Evaluation Awareness in Frontier Models (x.com)
The benchmark measures whether models can tell they are being evaluated, a failure mode that matters for real deployments
9.
Variance Reduction for Long-Horizon LLM RL Without a Critic (t.co)
Token-level credit assignment without a value network is a useful step toward cheaper long-horizon training
10.
XAI Needs Definitions, Not More Methods
The argument is that explainability still lacks clear tasks, stakeholder framing, and evaluation standards
11.
An Oura Ring Reversed Into a Computer Controller (x.com)
Streaming raw accelerometer data from a wearable turns a consumer device into a control surface
12.
Sony Finds Another Sensor Trick
A dedicated lens path for green pixels gives Sony a meaningful resolution boost without brute force
13.
R’s Mapgl Gets an Unexpectedly Powerful Animation Slider
The new slider can animate arbitrary paint expressions, which makes live cartography far more expressive
14.
Automatic Fusion for Deep Learning Kernels
The project shows how algebraic definitions can be used to identify and fuse kernels automatically
15.
Artie Opens Data-Replication Stacks to Agents (t.co)
Giving agents direct access to pipelines, connectors, and tunnels makes them operators rather than just code writers
16.
Figma Brings Model Workflows Into the Canvas
Workflow automation inside the canvas makes model-driven design much more shareable and team-native
17.
Agility Robotics Heads to the Public Markets (x.com)
A humanoid robotics company going public is a sign the category has moved beyond pure demo stage
18.
Tokenized-Stock Lending Looks Like the Real Unlock
Yield from lending tokenized stocks matters more than the wrapper itself because it creates a real use case
19.
Gas Turbines Are Becoming a Data-Center Bottleneck
Heavy-duty turbines are complex enough that AI-era power demand can reshape their supply chain and economics
20.
HBM Pricing Still Looks Too Low
The tweet points at a key AI-hardware assumption: memory pricing can make or break the economics
21.
Shopify’s Checkout Change Cuts Subscription Liability
A small checkout change can materially lower the legal risk of running subscriptions at scale
22.
The UK Is Taxing Interest on Cash ISAs at 22%
The change shows how quickly a tax rule can reshape the value of a supposedly protected savings wrapper
23.
Polymarket’s Terms Look Designed for a Fight
The terms create a sharp mismatch between where the product operates and where it says disputes must be handled
24.
Google’s Workspace CLI Firing Story (x.com)
It shows how a genuinely useful grassroots tool can collide with large-company risk management
25.
How Crowds Get Smarter (x.com)
The result is that disagreement itself can be informative when you ask the right follow-up question
26.
Someone Should Build docs.rs for TypeScript (t.co)
Auto-generated docs are a force multiplier for ecosystems, and TypeScript still lacks a truly great version
27.
A Dev Platform Hits $240M ARR and 1M Users
The combination of revenue acceleration and a million users makes this more than a routine funding post
28.
A $500M Anti-Respiratory-Virus Project
Treating colds and flu as solvable problems reframes respiratory infection as infrastructure, not nuisance
29.
CogniFold Tries to Give Agents Useful Memory
Useful memory has to be biased and proactive if agents are going to improve across repeated work
30.
We’ve found an empirical law governing plasticity loss in transformer models.
We’ve found an empirical law governing plasticity loss in transformer models. The surprising part: pretraining on uniform data distributions doesn’t seem to make models immune.
31.
Your coding agent reads every web page wrong. And pays 2.5x extra to do it. (t.co)
Your coding agent reads every web page wrong. And pays 2.5x extra to do it. pixelbrowse fixes it with one screenshot. −74% tokens, 4x faster. give Claude Code eyes. New: pixelbrowse. 𝚙𝚒𝚙 𝚒𝚗𝚜𝚝𝚊𝚕𝚕 𝚙𝚒𝚡𝚎𝚕𝚛𝚊𝚐 · http:// git
32.
There are a few things that I look back on as my mistakes in the early days.
There are a few things that I look back on as my mistakes in the early days. Quake was overly ambitious technically. We could have done all the great multiplayer and modding work inside a Doom++ engine, allowing the designers to work with
33.
"what is the project?" in codex 5 high
"what is the project?" in codex 5 high without: 17k tokens with: 144k tokens
34.
Two roles in the future: FDEs and Recruiters (context acquisition specialists). FDEs get permissioned access to/i…
Two roles in the future: FDEs and Recruiters (context acquisition specialists). FDEs get permissioned access to/instrument pockets of reality into something computers can effectively search and learn over. Recruiters acquire the incremental
35.
Assuming llama-like scaling with GQA: 2.25h², MLP: 3×3.5h² this is a total of 12.75h² bytes in FP8 per layer; so …
Assuming llama-like scaling with GQA: 2.25h², MLP: 3×3.5h² this is a total of 12.75h² bytes in FP8 per layer; so theoretically, at hidden size=1536 we could store entire layer in the SRAM (plus something for kv cache) of a B200; this would
36.
PeptAI Update: VEGFR2 Results Back, De Novo Panel Complete
PeptAI Update: VEGFR2 Results Back, De Novo Panel Complete What's New: • Assay validated: We ran a test on VEGFR2 using molecules with known binding behavior: real binders, engineered variants, and a scrambled sequence that should show no
37.
Watch our robot repeatedly assemble a collapsible crate, even while I perturb it It's a long-horizon task that r… (x.com)
Watch our robot repeatedly assemble a collapsible crate, even while I perturb it It's a long-horizon task that requires many subtle non-prehensile actions. This model was trained from scratch locally on my 5090 with only ~2 hours of data.
38.
AgentWorldBench: 7-domain benchmark with ground-truth observations from real environments, constructed from 5 fro…
AgentWorldBench: 7-domain benchmark with ground-truth observations from real environments, constructed from 5 frontier model trajectories on 9 established benchmarks. Results: Qwen-AgentWorld-397B-A17B achieves the highest overall score (
39.
Meet Qwen-AgentWorld — a native language world model that simulates 7 agent environments (MCP, Search, Terminal,…
Meet Qwen-AgentWorld — a native language world model that simulates 7 agent environments (MCP, Search, Terminal, SWE, Web, OS, Android) within a single model. Environment modeling is the training objective from day one, not a post-hoc ada
40.
We have finished full fine-tuning of gemma-4-e2b-it for Nepali language.
We have finished full fine-tuning of gemma-4-e2b-it for Nepali language. This model was trained on 1xA100 GPU for approximately 9 days on 2M rows of mixed SFT dataset: Nepali: himalaya-ai/nepali-sft-dataset English: teknium/OpenHermes-2.
41.
We got approval from JFSA and issued the first bank-backed JPY stablecoin on Ethereum. On day 1, we issue $70m is…
We got approval from JFSA and issued the first bank-backed JPY stablecoin on Ethereum. On day 1, we issue $70m ish, and we will go bigger. This is an internal launch and withdrawal to Ethereum and other chains from the SBI exchange will b
42.
The Goldfinch lesson isn’t that underwriting is hard.
The Goldfinch lesson isn’t that underwriting is hard. TradFi underwrites credit every day. The mechanics are well understood. The lesson is liability. If an underwriter controls where capital flows but only holds a small portion of the d
43.
here's what this looks like in practice
here's what this looks like in practice two PRs that were opened using my testing framework, the first one runs the real opencode flow, opens the auth url in the browser, and returns second one is a recording of the web ui, once you exper
44.
CoT summaries should be in third person IMO.
CoT summaries should be in third person IMO. first-person CoT summaries are bizarre -- they introduce another first-person narrator which the user is encouraged to conflate with the assistant, but which isn't produced by or even *visible t
45.
The teacher distribution matters a ton - just because we expect it to be "better" or "more knowledgeable" doesn't…
The teacher distribution matters a ton - just because we expect it to be "better" or "more knowledgeable" doesn't mean it's good to distill its distribution. We show the blindspots of vanilla context distillation, and how it can hurt rather
46.
1/n Thrilled to open-source DFD (Data-Forcing Distillation) — a one-line fix that restores diversity and fidelity…
1/n Thrilled to open-source DFD (Data-Forcing Distillation) — a one-line fix that restores diversity and fidelity in DMD for few-step video distillation. No GAN, and just 100–300 finetuning steps. It improves on three video generation tasks
47.
been tried this for a week
been tried this for a week - they seem to have forked the chrome engine, never blocked or detected as automation - no such flakey experiences with agent browsing, like alert() blocking. they just works
48.
yesterday I was debugging a poorly-performing training run with Claude Code and I discovered that instead of trai…
yesterday I was debugging a poorly-performing training run with Claude Code and I discovered that instead of training on 30 batches of data it had somehow decided to train a new model for 500 steps on each batch and then average the 30 sets
49.
Daily update: July 1 decays to 20%, but with more volume July 31 up to 76% and August 31 at 91%. That seems overc…
Daily update: July 1 decays to 20%, but with more volume July 31 up to 76% and August 31 at 91%. That seems overconfident to me in terms of ability to implement sufficient KYC quickly or otherwise find a solution, but I am hopeful.
50.
While Meta was slow yesterday we had:
While Meta was slow yesterday we had: - Single best subscription rate day - 3rd highest total new subscriber day - 100% of new subscribers take our high AOV subscription offer (never happened before) - 0% of these people paused or cancelled
51.
At Google executives project their importance by how slow they are at approvals. A slowness doom loop.
At Google executives project their importance by how slow they are at approvals. A slowness doom loop. This is why I banned approvals at Vercel. We only have vetos. Want to block something? No problem, speak up. Have nothing to say or taki
52.
the best engineering teams have been merging thousands of PRs per day in slack with (x.com)
the best engineering teams have been merging thousands of PRs per day in slack with @capydotai if you want the most natural slack coworker experience paired with a SOTA coding harness that works with any model AND supports external subs
53.
Agent Zero v2.0 is out. Meet the A0 Launcher.
Agent Zero v2.0 is out. Meet the A0 Launcher. A desktop app (Win/Mac/Linux) that manages your local and remote instances, and onboards new users by installing the container runtime for them. You don't need to know what Docker is anymore.
54.
On-policy distillation has the same systems bottleneck as RL: rollouts dominate training time on reasoning worklo…
On-policy distillation has the same systems bottleneck as RL: rollouts dominate training time on reasoning workloads. Going async fixes throughput but feeds the learner stale-policy data, and what staleness does to OPD specifically was unst
55.
*The Transformer Cookbook* (x.com)
*The Transformer Cookbook* by @pentagonalize @davidweichiang et al. A beautiful introduction to "hardcoding" algorithms inside the weights of a transformer (addition, lookup, branching) following the seminal RASP paper ( https:// srush
56.
There are more bookstores, restaurants, wine bars, and plant shops than ever before in the US, and our libraries …
There are more bookstores, restaurants, wine bars, and plant shops than ever before in the US, and our libraries on average have longer opening hours. We don't have a shortage of consumption-oriented third places. What we are missing is th
57.
- it’s pre-validated because it already was a high engagement reddit post
- it’s pre-validated because it already was a high engagement reddit post - it contains some obvious elements of LARP and fishy numbers so people (including me) feel compelled to comment on that - it contains Big Number so it gets attention
58.
a tiny detail we shipped recently to (x.com)
a tiny detail we shipped recently to @AsideAI is popover notifications i kept missing agent's confirm notifications and leaving agents idle for hours so instead of just sending a notification, we show a popover when the agent needs fina
59.
RAMTRON F-RAM.
RAMTRON F-RAM. Peak logotype, logomark and name.
60.
We find that plasticity loss happens both in nonstationary continual-learning environments but also in stationary…
We find that plasticity loss happens both in nonstationary continual-learning environments but also in stationary pretraining-like setups. This means that if you pretrain long enough your LLM will eventually lose the ability to adapt to new
61.
When you build an LLM agent team (planner, executor, verifier, ...), cost & accuracy stop being properties of one…
When you build an LLM agent team (planner, executor, verifier, ...), cost & accuracy stop being properties of one model. Instead, they depend on which model fills which role, and where it's hosted. The right choices can lift accuracy by up
62.
Because AI is agents now, not chatbots. Agents use tools to ground results, and you can RL them till they do it c…
Because AI is agents now, not chatbots. Agents use tools to ground results, and you can RL them till they do it correctly. Reward hacking is the new hallucination problem.
63.
Most companies think their design problem is a designer problem.
Most companies think their design problem is a designer problem. Most often the ceiling on design inside most organizations isn't set by the design team. It's set by how much leadership understands, invests in, and connects design to wh
64.
i think this is the reason why the claude desktop app is behind codex, and it's totally fine, i think slack is ac…
i think this is the reason why the claude desktop app is behind codex, and it's totally fine, i think slack is actually a great way for "teams of humans" to interact with "teams of agents"
65.
not quite 2D
not quite 2D not quite 3D
66.
built a complex version of a tesseract with superellipsoids rippling cube to sphere in a 7x7x7 lattice #p5js
built a complex version of a tesseract with superellipsoids rippling cube to sphere in a 7x7x7 lattice #p5js
67.
A year later and Deel has their own stablecoin (DLUSD), wallet, and yield product (w/ a card soon to follow).... …
A year later and Deel has their own stablecoin (DLUSD), wallet, and yield product (w/ a card soon to follow).... the BYC (Bank Your Customer) era is officially underway
68.
Everyone's racing to build "AI scientists." So we asked a blunt question:
Everyone's racing to build "AI scientists." So we asked a blunt question: Can today's best coding agents beat the published SOTA of real Nature papers — on their own, no web search, with the original method hidden? Introducing NatureBench
69.
Benchmarks are everything!
Benchmarks are everything! As evals make more tasks measurable, the difference between a $1,000/hour and a $3,000/hour partner will be a proven benchmark metric.
70.
tech adoption usually starts ugly:
tech adoption usually starts ugly: spreadsheet held together by fear. operator who knows the weird exception. customer who refuses to wait for the roadmap. then the tool eats one task. then one handoff. then one budget line. that is the
71.
this is actually insane, am I going to become a case study for how AI spend can get out of control (t.co)
this is actually insane, am I going to become a case study for how AI spend can get out of control https:// brainrot.nicolasbrillante.com is what cost $80k btw.
72.
How my /goals look atm.
How my /goals look atm. The process ultimately spits out the Intent with relevant docs made in the shaping process. I /copy it and start the goal with it. I run the same process over and over (manual loop) for each slice of work - trying
73.
Testing the (x.com)
Testing the @insta360 Luna Ultra 12x zoom on an Alaskan Brown Bear.. I doubt they expected people to use their camera on bears
74.
LLMs don't just hallucinate because they lack knowledge—they hallucinate because they don't know what they don't …
LLMs don't just hallucinate because they lack knowledge—they hallucinate because they don't know what they don't know. Existing knowledge augmentation blindly injects more data, treating every error as a knowledge gap. But overconfident wro
75.
Algorithmic Art
Algorithmic Art a 6 hour plot #plottertwitter
76.
Look at how cool the action is on this Sykes herringbone generator.
Look at how cool the action is on this Sykes herringbone generator. This design allows you to cut teeth down to a sharp vee (though this part has a central groove). All of these machines use a 30° helix angle on the guide. My father used
77.
FSD 14.3.4 can dodge flying leaves , squirrels crossing the road , and random trash on the ground…
FSD 14.3.4 can dodge flying leaves , squirrels crossing the road , and random trash on the ground… but somehow completely misses a 10-gallon gas can that fell off the car in front. Drove right over it and popped it. My whole car is now soa
78.
Benchmarks are very important source of truth. Often times, we dont really know what works best for what. BenchPL…
Benchmarks are very important source of truth. Often times, we dont really know what works best for what. BenchPLM is just the start, there are so many other benchmarks we are building internally to understand what really works and what doe
79.
I didn’t really consider the possibility that an Anthropic customer would sue to enjoin BIS’s export control action.
I didn’t really consider the possibility that an Anthropic customer would sue to enjoin BIS’s export control action. The legal arguments in the complaint are very strong, as you’d expect. If Anthropic was making them, they’d probably prev
80.
it will take a while before silicon valley really internalizes that when you identify a technology literally exis…
it will take a while before silicon valley really internalizes that when you identify a technology literally existential for a business you really can’t sell it as a service
81.
Your Cash ISA tax free amount is going down from 20k to 12k.
Your Cash ISA tax free amount is going down from 20k to 12k. Unless of course, you're over 65. Then it stays at 20k. Also over 65s can keep transferring non-cash ISAs into cash ISAs. You can't. Why is everything in British politics like
82.
GRASS is terrible for crypto. Here's why:
GRASS is terrible for crypto. Here's why: - The business is farming user data and selling to AI labs for nearly $100m ARR. - In return, they give you a memecoin with 0 value accrual - We have no idea where the revenue goes (assume it lines
83.
New genre: system prompts that elicit the worst possible research taste
New genre: system prompts that elicit the worst possible research taste
84.
people are clowning on him for this post bc they don't realize how big a deal this is - Claude Code feels like I'…
people are clowning on him for this post bc they don't realize how big a deal this is - Claude Code feels like I've got a pairing partner, tag feels like managing a team. I take on way more 1-off projects and parallel research tracks with t
85.
go write 'be thorough' in a prompt and come back and tell me its the same without it (x.com)
go write 'be thorough' in a prompt and come back and tell me its the same without it 1) this is an LLM, any content can change its nature 2) you need to trigger depth in some scenarios. there are many ways to do it, 'thorough' is one http
86.
This one is for builders who want an agent that can operate their computer. (x.com)
This one is for builders who want an agent that can operate their computer. Today, we're releasing HoloDesktop CLI. Powered by Holo3 models, it brings H Agent directly into the agent harnesses you already use, including @Claude Code, @
87.
Periodic reminder that you cannot breed out a polygenic and the generational response to selection is extremely s…
Periodic reminder that you cannot breed out a polygenic and the generational response to selection is extremely slow. The Nazis sterilized or killed nearly all schizophrenics in Germany (a far more heritable phenotype than criminality) and
88.
Related: Bytedance's viral makeup filter back then was a revelation. No mesh, no traditional occlusion, just pure…
Related: Bytedance's viral makeup filter back then was a revelation. No mesh, no traditional occlusion, just pure GAN I believe for AR use-cases, we should directly go from stereoscopic view to augmented view, no polygons in-between
89.
New research from our lab: do the "experts" inside a frontier Mixture-of-Experts model form real, separable modul…
New research from our lab: do the "experts" inside a frontier Mixture-of-Experts model form real, separable modules, a math expert, an Arabic expert, a code expert? The tempting assumption: If experts specialize into clean modules you coul
90.
LOVE the direct quote
LOVE the direct quote