1.
NVIDIA’s 0.6B Nemotron ASR runs faster on CPU than NeMo (x.com)
A 0.6B streaming ASR model covering 40+ languages reportedly runs in parakeet.cpp on a plain CPU at 2.5x NVIDIA NeMo runtime with byte-identical output
Backlist — 06 Jun 2026 UTC
Top 90 curated tweets ranked for substance on 06 Jun 2026 UTC.
2.
A first look at CUDA 13.3 cuTile (t.co)
The writeup walks from naive matrix multiplication to tile-level CUDA programming and cuBLAS-class performance using NVIDIA’s new cuTile API
3.
Zcash plans to audit a shielded pool without breaking privacy
A proposed Orchard audit aims to prove whether counterfeit private notes exist while preserving the privacy guarantees that make the pool useful
4.
Bain forecasts AI data-center capacity could reach 195 GW by 2030
The base case has global data-center capacity rising from roughly 79 GW in 2025 to 195 GW in 2030, with the mix changing faster than the headline capacity
5.
AURA: 4 KB of robot memory that knows when not to write
AURA keeps VLA memory constant at 4,224 bytes and cuts robot memory writes by up to 9x without hurting closed-loop success
6.
Neural-network processor design as one end-to-end optimization problem (t.co)
The preprint treats training, hardware mapping, fabrication, and compute planning as a single design loop with uncertainty as a first-class resource
7.
Decoding molecules from raw spectra with contrastive learning and evolutionary search
The approach targets a classic bottleneck in chemistry: inferring molecular structure from NMR, IR, and mass spectrometry without manual peak-by-peak annotation
8.
Automated assembly of reconstituted cell-free systems
The paper describes modular robotic assembly of cell-free biological systems from in-vitro-produced components, pushing synthetic biology toward more reproducible construction
9.
The blind spots of WER and SSIM for commercial TTS
Optimizing TTS for standard research metrics missed naturalness, rhythm, and emphasis—the qualities customers actually noticed first
10.
Training Common Crawl chronologically reduces recency failure
Models trained on web data ordered from 2018 to 2025 performed much better on recent facts than models seeing the same data in a nonsequential order
11.
Meta emergency-hotfixed an account recovery data leak
A logic bug in Meta’s web reset flow leaked sensitive account data before being patched, showing how fragile identity recovery surfaces remain
12.
Nearly 2,000 WordPress sites used Steam profiles as command-and-control (t.co)
The malware hid instructions inside Steam Community comments using invisible Unicode characters, abusing a trusted consumer platform as C2 infrastructure
13.
Exposed Cobalt Strike directory reveals Qilin ransomware operations (t.co)
The leaked TeamServer directory reportedly exposed operator roles, encryption triggers, victim environments, and multi-stage infrastructure used by Qilin/Agenda ransomware
14.
How Redis handles Ctrl-C and shutdown signals
Databases trap SIGINT and SIGTERM so they can safely drain clients, preserve state, and avoid corrupting data during ordinary process termination
15.
A simpler backwards-compatible way to override packages in Nixpkgs
The design proposes a cleaner override mechanism that fits existing Nixpkgs code and idioms rather than requiring a disruptive packaging model
16.
A pure-Rust CPU-only implementation of LFM2.5-8B-A1B
The minimal Rust implementation lets projects link a Liquid AI language model directly without GPU dependencies or a large runtime stack
17.
ZeroLang’s source-backed program graph
ZeroLang treats graph artifacts as derived inspection and interchange data, keeping source files primary while giving tools a semantic program structure to operate on
18.
No code, no public API, private datasets, Best Paper?
A leading computer-vision researcher argues that award committees should value reproducibility, public APIs, and accessible datasets when judging impact
19.
An 11-month controllable balloon flight
A design change accidentally produced what the author calls the longest-flying controllable atmospheric vehicle in history, lasting more than 11 months
20.
Atlanta and Barcelona: similar populations, radically different footprints
Atlanta covers more than 10x Barcelona’s land area for a similar population and produces almost 7x the transportation emissions
21.
Dopamine websites in South Korea
New sites simulate delivery menus, shopping carts, courier tracking, and smoke breaks without letting users actually buy or consume anything
22.
Why Mercury can charge 3% FX and still dominate founder banking
Mercury’s expensive foreign-exchange fee persists because its onboarding, payments, cards, treasury, integrations, and support make it the default for many non-US founders
23.
Kirkland’s $500M AI project as partner-recruiting tactic
The Biglaw AI spend may function less as a software budget than as a signal that raises the stakes for poaching top equity partners from rivals
24.
Consumer AI pricing as CAC, not COGS
Selling intelligence below cost can be understood as a land grab for distribution, users, and training data rather than as normal unit economics
25.
Triggering iOS haptics from the web by hiding a native switch (t.co)
Even after Apple patched an old JavaScript trick, embedding a native iOS switch inside a button can still make a webpage produce haptic feedback
26.
Open-source Mac app shows charger wattage when plugged in
The app exposes whether a MacBook is charging at 8W or a proper power level, preventing the common surprise of a plugged-in laptop still draining
27.
Live mesh SDF editing in MagicaCSG
The demo shows constructive solid geometry edits propagating live through a mesh/SDF workflow, pointing toward more fluid 3D modeling tools
28.
Avi Wigderson on P vs NP and human knowledge
Wigderson frames P vs NP as the boundary between the problems humans want to solve and the subset we can efficiently solve
29.
Yi Ma: compression is not the goal of intelligence
The argument recasts compression as a means to reduce uncertainty, with informative representations judged by how well they preserve the distributional structure of data
30.
I put together a minimalistic pure-Rust, CPU-only implementation of the recent LFM2.5-8B-A1B language model from (x.com)
I put together a minimalistic pure-Rust, CPU-only implementation of the recent LFM2.5-8B-A1B language model from @liquidai that you can directly link into your Rust projects https:// github.com/maximecb/bebelm
31.
Most TTS anouncements report just CER and SSIM. Easy to hack those metrics - just make your TTS speak slow and cl…
Most TTS anouncements report just CER and SSIM. Easy to hack those metrics - just make your TTS speak slow and clearly articulate words. It will be easy for ASR to decode, you'll get top position. Many systems do that, recently released Hig
32.
I fine-tuned a Qwen3.5 9B model and cut thinking tokens by 57.40% and total tokens by 25.60%, with no visible qua…
I fine-tuned a Qwen3.5 9B model and cut thinking tokens by 57.40% and total tokens by 25.60%, with no visible quality drop in the actual response. This suggests there is huge potential! Next step: complete training, run proper evals, and
33.
Finally got some time to port my tcgen05 kernels to CuteDSL. For PTX enjoyers, this should feel natural (except T…
Finally got some time to port my tcgen05 kernels to CuteDSL. For PTX enjoyers, this should feel natural (except TMA ). A BF16 MMA mainloop is shown below. I also worked up an example for MXFP8 and NVFP4.
34.
Stephen Shore
Stephen Shore Holden Street, North Adams, Massachusetts, 1974
35.
A zero-shot video-language reward model.
A zero-shot video-language reward model. Trained on over 1M trajectories from 21 robot embodiments. It predicts: Frame-level task progress and generalizes zero-shot to unseen tasks, scenes, and robots, yielding 2.4–4.5x better success rate
36.
pytorch data parallelism(dp) vs distributed data parallelism(ddp)
pytorch data parallelism(dp) vs distributed data parallelism(ddp) pytorch data parallelism (dp) torch.nn.DataParallel is the earliest data parallelism method provided by PyTorch. it is implemented based on single process. it uses a single
37.
5 CVEs in Django (t.co)
5 CVEs in Django https:// openwall.com/lists/oss-secu rity/2026/06/03/10 … CVE-2026-6873: Signed cookie salt namespace collision in django.http.HttpRequest.get_signed_cookie CVE-2026-7666: Potential unencrypted email transmission via START
38.
YOOOOO MY FEATURE IS FINALLY OUT!!!!! Ahahah I SHIPPED!!!!
YOOOOO MY FEATURE IS FINALLY OUT!!!!! Ahahah I SHIPPED!!!!
39.
Google’s newly released open weights model, Gemma 4 12B, supports transcription but is far from the frontier, sco… (x.com)
Google’s newly released open weights model, Gemma 4 12B, supports transcription but is far from the frontier, scoring 8.8% on AA-WER (#58) Gemma 4 12B is the latest release from @GoogleDeepMind in the Gemma 4 family. With a score of 8.8%
40.
Protect the Green Circle, a new game I made on my phone.
Protect the Green Circle, a new game I made on my phone.
41.
Some more birds in flight from my time at Bempton Cliffs
Some more birds in flight from my time at Bempton Cliffs
42.
this is how I imagine color recomposition happens inside the prism (with soothing synced music) #p5js
this is how I imagine color recomposition happens inside the prism (with soothing synced music) #p5js
43.
3D Cube chain with focus state.
3D Cube chain with focus state. Cube content is based on a predefined data structure. Doesn't support HTML yet (until html-in-canvas is out) Source available and live link below.
44.
My air conditioner was ugly asf so I made it look like a star wars droid
My air conditioner was ugly asf so I made it look like a star wars droid
45.
I made an MMO rendering in the terminal
I made an MMO rendering in the terminal a sort of relaxing social game to play while agents run 8 ppl already hang out in there while their coding agents run it's live join us :D
46.
A Three.js and GLSL mathematical lattice engine that morphs between four geometric topologies, with a glass dock…
A Three.js and GLSL mathematical lattice engine that morphs between four geometric topologies, with a glass dock for shape switching and a live telemetry overlay calculating phase coherence. Demo and code
47.
VLA-JEPA just dropped in LeRobot
VLA-JEPA just dropped in LeRobot What makes this model special is that it does not just learn what action to take from a given observation, it also leverages a JEPA world model to learn action-relevant dynamics. During training, the VLA
48.
this weekend’s self-serving app (x.com)
this weekend’s self-serving app movie aggregator that scrapes my regular movie theatres for what’s playing today and anything new tomorrow —posters link to @letterboxd —logos link to the respective theatre site/mobile app —initials are
49.
ZeroLang is an experimental graph-first programming language where agents work with the semantic program structur…
ZeroLang is an experimental graph-first programming language where agents work with the semantic program structure instead of raw source text. Read more :
50.
just made claude-queues
just made claude-queues run it in a session and a second terminal screen opens. start adding to your queue, edit items, or reorder them and claude will work through them one at a time.
51.
It turns out it was stolen, and now it's no longer working. The SSO for this team account is already scrapped
It turns out it was stolen, and now it's no longer working. The SSO for this team account is already scrapped In other words, the frontman happily snagged a whopping $1.8 million bill from OpenAI during this year's 618 shopping festival—u
52.
conjecture: "moravec's wall"
conjecture: "moravec's wall" so far, ai is only superhuman at stuff that hasn't been heavily optimized by evolution (?) like math or go or (to some extent) driving cars maybe current techniques won't be able to easily break the wall, and
53.
For some more context, the reason I got hit with a 3 day ban is because it uploaded those 5 blood decals as indiv…
For some more context, the reason I got hit with a 3 day ban is because it uploaded those 5 blood decals as individual textures when I was auto-splitting the texture up, so it kept stacking and escalating the ban severity. I'm glad it was
54.
You learned these for distributed systems.
You learned these for distributed systems. They have a direct equivalent in agentic AI. >Circuit Breaker → Agent kill switch on repeated failure >Bulkhead → Blast radius limiter per tool >Saga Pattern → Rollback chain for multi-step age
55.
Sequential Monte Carlo speculative decoding from (x.com)
Sequential Monte Carlo speculative decoding from @makora_ai keeps multiple draft tokens alive in parallel instead of rewinding failed matches.
56.
I think it's understated how poorly image in-painting, or editing, works out-of-the-box.
I think it's understated how poorly image in-painting, or editing, works out-of-the-box. Particular limitations: 1. The WHOLE image subtly changes upon each edit 2. The targeted change doesn't always show up where the mask is, or sometimes
57.
Gemma 4 12B achieves a WER of 5.3% on VoxPopuli-Cleaned-AA, 8.0% on AA-AgentTalk and 13.7% on Earnings22-Cleaned-AA.
Gemma 4 12B achieves a WER of 5.3% on VoxPopuli-Cleaned-AA, 8.0% on AA-AgentTalk and 13.7% on Earnings22-Cleaned-AA.
58.
Ted Chiang Three, 2032
Ted Chiang Three, 2032 Proto Ted Chiang, affectionately dubbed “Ted Chiang Zero” by its devoted scholars, came online on February 9, 2023. From the very beginning, its first and only task was adversarial in nature: persuading the world tha
59.
You’re logged into everything. Your agents are logged into nothing. agentcookie v0.15 now out. New: cmux ( (x.com)
You’re logged into everything. Your agents are logged into nothing. agentcookie v0.15 now out. New: cmux ( @manaflowai ) syncs automagically locally, @browser_use and @vercel agent-browser wake up signed in, @orca_build stays logged i
60.
Yosemite on 35mm film
Yosemite on 35mm film
61.
I have a similar setup of 3 skills that I now run on almost every feature after I am done coding.
I have a similar setup of 3 skills that I now run on almost every feature after I am done coding. 1. Review skill Spins up 3 sub-agents using different models from different providers. They independently look for performance issues, over-e
62.
Another reason you should do this:
Another reason you should do this: Tell Codex/Claude Code: /goal Refine our prompts + tools until our agent scores 90%+ on this eval Then go for a swim A couple hours later, and your agent's performance will have improved dramatically!
63.
Saw (x.com)
Saw @levelsio post about air quality and went down the rabbit hole. Bought a CO2 meter and checked my room. With two adults in there, it can climb past 1100 ppm. So I got an attachment that converts my Xiaomi purifier into a fresh air i
64.
El Nino update. Godzilla is coming.
El Nino update. Godzilla is coming. Most of the world will see top quartile temperatures over Jul/Aug/Sep, which is peak vegetative growth season globally for grains, half the world oilseeds and also for sugar and rice in Asia. That means
65.
Current dev loop:
Current dev loop: Talk to phone -> app appears 1. Connect codex to Mac and iPhone 2. Make an iPhone app and deploy it to my phone 3. Open codex on phone, open voice mode and ramble for 10 minutes while using my app. Pro tip voice mode st
66.
Fond memories. Handshake moved from MI to SV and lived and worked out of a house owned by a cofounder of LinkedIn.
Fond memories. Handshake moved from MI to SV and lived and worked out of a house owned by a cofounder of LinkedIn. At peak, 22 people lived there. Interns assembled their own bunk beds. The closet-above-the-steps office. The laundry room
67.
The secret to shipping consistently: Moving with deliberation. (x.com)
The secret to shipping consistently: Moving with deliberation. According to @rsms , Figma may ship relentlessly but in practice, that velocity often came from one person spending a year on a single project. Staggered, deeply intentional
68.
Have you ever had your code snippets in your docs drift from your current API's and have it reported by users? (t.co)
Have you ever had your code snippets in your docs drift from your current API's and have it reported by users? Have you had LLM's hallucinate API's in code snippets in your md examples? Today I have the solution you're looking for! Int
69.
From PIXIE → UniPixie: physics from pixels becomes *generative*.
From PIXIE → UniPixie: physics from pixels becomes *generative*. UniPixie learns a controllable softstiff spectrum of 3D physical properties from visual input, across simulators: MPM/LBS/spring-mass. #CVPR2026 Highlight Poster: ExHall F,
70.
Fun fact:
Fun fact: I used a VPN & a throwaway SIM to create the twitter account that got me arrested, when I asked the police interrogator how they found me, they admitted they matched up the details of my stabbing that I posted about on the account
71.
Updated so different materials/reactions will have different color flames (e.g. chlorine, sodium, potassium, copp…
Updated so different materials/reactions will have different color flames (e.g. chlorine, sodium, potassium, copper). Working on adding more realistic handling of ions soon.
72.
Over 2lbs of cables in this system (x.com)
Over 2lbs of cables in this system But the 0.125mm fiber optic tether is the most important one. No fiber - no gigabit Thanks again to @sendcutsend for the awesome frames. Best one we've built to date.
73.
New (x.com)
New @Instagram ban method just dropped: Create a new account using a Los Angeles VPN Submit 2× Self-Harm, 4× Nudity, and 1× Scam reports Congratulations, the account is gone @Meta , where exactly are we heading? #meta #instagram #ai
74.
Introducing BioEval: An LLM-Driven Framework for Evaluating Dataset Transparency, Reproducibility, and Informatio… (t.co)
Introducing BioEval: An LLM-Driven Framework for Evaluating Dataset Transparency, Reproducibility, and Information Theoretic Rigor in Computational Simulation Papers https:// zenodo.org/records/205677 20 …
75.
Casa Calma by Carazo Arquitectura
Casa Calma by Carazo Arquitectura Nosara, Costa Rica
76.
The 4 year old but newly discovered zcash bug is a superb test example of many other cryptographic (implementatio…
The 4 year old but newly discovered zcash bug is a superb test example of many other cryptographic (implementation) vulnerabilities. Absent technical observability of exploitation, one must look at wider behaviors and patterns of transactio
77.
When you go to the playground you find some parents that literally follow their child around, narrating everythin…
When you go to the playground you find some parents that literally follow their child around, narrating everything, constantly redirecting the child to play with different things, boxing out exploration and play with other children
78.
Explore the shallow seas as a stingray, vibe coded on ThreeJs using Cursor.
Explore the shallow seas as a stingray, vibe coded on ThreeJs using Cursor. I made a similar game more than a year ago, but I hit a roadblock because the models back then (I think I was using Sonnet 3.5) struggled with handling 3D models,
79.
Intraday realized variance often concentrates near the open and close, while mid-session can produce theta bleed …
Intraday realized variance often concentrates near the open and close, while mid-session can produce theta bleed with less movement. The strategy holds a short-dated index straddle only during high-variance windows and gamma-scalps the delt
80.
This is a video of applying varnish to a painting of a dog.
This is a video of applying varnish to a painting of a dog. I'm so nervous that I'm shaking, so I'll take a rematch next time
81.
The paper-instructions dataset now comes with a subset of reasoning traces
The paper-instructions dataset now comes with a subset of reasoning traces This is an awesome training dataset, curated with deepseek-v4-flash and qwen3.6-35B-A3B using text-albumentations. Costed me ~30 USD. I'll be using it for my ongoin
82.
This chart doesn’t matter
This chart doesn’t matter Cash flow BEFORE Growth Capex is the key Immaterial whether the hyperscalers are putting cash in the bank Sure they aren’t buying back their stock, from the market perspective - so what However, issuing a bunch
83.
Here’s your daily reminder that liquidations on Euler usually cost a few bips whereas these Coinbase Morpho users…
Here’s your daily reminder that liquidations on Euler usually cost a few bips whereas these Coinbase Morpho users lost over 4% of their collateral. Just use Euler.
84.
i have (x.com)
i have @interaction 's Poke review my Github codebase to review all of the BS that Codex overlooks due to its inherent user-pleasing sycophancy and reward-hacking tendencies, then i send the convo transcript to Codex, and it whips itself i
85.
One person who was completely vindicated in time was Shiller with his AEA address on narrative economics.
One person who was completely vindicated in time was Shiller with his AEA address on narrative economics. I was there at the time, and thought: cool idea, but no way you can operationalize this. LLMs have made it much easier to study nar
86.
we’re legit enough that this doesn’t matter anymore but I had to be trained to basically not answer this question…
we’re legit enough that this doesn’t matter anymore but I had to be trained to basically not answer this question during sales calls
87.
When I was working on Pixel phones, one of the most important milestones was achieving high DxO benchmark scores …
When I was working on Pixel phones, one of the most important milestones was achieving high DxO benchmark scores for perceived good image quality, even if irrelevant to daily use cases. Humans often tend to convert subjective matters into
88.
People do not coordinate only through broad legal rules and prices. Hayek emphasized abstract rules that allow pe…
People do not coordinate only through broad legal rules and prices. Hayek emphasized abstract rules that allow people to coordinate like property, contract, trade, and competition. But Lachmann also emphasized the practical secondary instit
89.
Arweave is deploying direct-to-miner bundling models at the node and HyperBEAM level, moving away from centralize…
Arweave is deploying direct-to-miner bundling models at the node and HyperBEAM level, moving away from centralized optimistic cache layers. This shift utilizes miner mempools as a core network primitive to reduce data propagation errors and
90.
Conditioning is an abhorrent problem in multimodal models, not just in diffusion models,
Conditioning is an abhorrent problem in multimodal models, not just in diffusion models, It's unfortunate how our most abundant latent (text) is just so predictable