1.
What a 110k-GPU cloud deal implies about GB200 lease rates (t.co)
The back-of-the-envelope puts a $920M/month, 110k-GPU agreement at roughly $11.50 per GPU-hour, a concrete datapoint for frontier AI compute economics
Backlist — 07 Jun 2026 UTC
Top 90 curated tweets ranked for substance on 07 Jun 2026 UTC.
2.
Agent firewalls are the wrong abstraction
Dangerous agent actions are better constrained with permissions and sandboxing than inferred after the fact by an AI runtime security layer
3.
Replacing string-interpolated codegen with a table and an interpreter
Moving from generated code to generated data erased escaping bugs and cut a system by about ten thousand lines
4.
Folddisco: search millions of protein structures for a 3D motif in seconds (x.com)
Fast 3D motif search across millions of structures changes both enzyme design validation and annotation of uncharacterized proteins
5.
Stop using ipTM for AlphaFold PPI screens
ipTM can mislead AlphaFold interaction screens when irrelevant IDRs or domains skew the score, making ipSAE or LIS safer for high-throughput PPI work
6.
US data-center construction spending now exceeds transportation structures
Census construction data puts data-center spending at a $50.7B annual rate, above public transportation-structure spending
7.
3D DRAM and co-designing memory with compute
Stacking DRAM on compute creates a new co-design space for memory bandwidth, power, and computer architecture
8.
A custom pnpm registry with 2–7x faster installs (x.com)
Package install time remains a major dev-loop bottleneck, and a registry-layer change claims multi-fold gains over pnpm's already-fast Rust path
9.
Storing files in DNS resolver caches
Open DNS resolver caches become a bizarre distributed storage medium when millions of resolvers can temporarily hold shards of a file
10.
Meta account recovery reportedly exposes PII from a username
Unauthenticated account recovery revealing emails or phone numbers from usernames would be a severe privacy failure at social-network scale
11.
C0XMO botnet variant exploits DD-WRT routers and kills rivals
The malware targets old DD-WRT routers across multiple CPU architectures, supports 19 DDoS methods, and removes competing implants
12.
Humanoid-GPT: zero-shot humanoid control from 2B motion frames (x.com)
A transformer trained on two billion motion frames reportedly transfers dynamic behaviors like soccer and digging to a real G1 robot without fine-tuning
13.
Robotics data collection is bottlenecked by manual ops
The hardest part of frontier robot datasets is often camera charging, sync, placement, and reliable upload rather than model training
14.
Require code and weights for CVPR best-paper eligibility (x.com)
Making awards depend on source and weight release would turn prestige into leverage for reproducible computer vision research
15.
AI training is stuck in a Frankenstein multi-stage local minimum
A frontier researcher argues that pretraining, post-training, RL, and tool scaffolding should give way to continual interaction and causal training
16.
Omni: local multimodal file search for macOS (t.co)
The Apache-licensed app indexes text, PDFs, images, audio, and video locally on macOS using multimodal embeddings
17.
elio: a batteries-included terminal file manager in Rust (x.com)
The Rust TUI combines image previews, fuzzy search, zoxide, trash support, shell integration, and themes into a single file-manager workflow
18.
A multi-warp SIMT GPU core in SystemVerilog
Implementing branch divergence with an IPDOM stack is a concrete step toward understanding real GPU execution rather than treating GPUs as magic
19.
Sniffnet: a human-readable network monitor with 37k GitHub stars
A simpler packet-monitoring UI lowers the barrier for people who need network visibility but bounce off Wireshark
20.
StableOrNot: test your intuition on protein stability prediction (t.co)
Turning model benchmarks into a playable stability-prediction game makes it easier to understand what protein ΔG predictors actually get right
21.
A real-data heatmap of San Francisco one-bedroom rents
A neighborhood-level rent map shows the distribution behind San Francisco's housing debate, including a reported SoMa rebound
22.
Dormant Bitcoin wallets, abandoned-property law, and proof of life
A lawsuit trying to claim 39,069 dormant wallets as abandoned property was answered by an on-chain signature from an old Bitcoin owner
23.
South Korea’s dopamine sites simulate shopping without purchase
Fake carts, fake delivery tracking, and simulated orders show how commerce UX can be detached from commerce itself
24.
Audacity now ships with launcher, cloud popups, and upsells
A famously lightweight open-source audio editor adding accounts and cloud prompts is a concrete example of software bloat reaching trusted tools
25.
Digital archives disappear faster than people expected
Digital information is not self-preserving, and the first decades of the web are already suffering link rot and archive loss
26.
Live chat as a sales tool for developer products
Developers often refuse calls but ask precise technical questions when blocked, making support chat a high-signal buying channel
27.
Why single-digit-billion-dollar drugs can be unattractive to big pharma
Approval costs can make even billion-dollar markets too small for large pharma portfolios, shaping which diseases get funded
28.
Zcash Ironwood proposes a formally verified shielded pool
A new shielded pool backed by formal verification and audits would be an unusually rigorous upgrade path for privacy-preserving cryptocurrency
29.
Collision-aware tooltip positioning in CSS/JS
Correct tooltip behavior needs viewport measurement, dynamic placement, arrow movement, and animation origins rather than a fixed absolute position
30.
Cardbal: read transit-card balances and routes on iOS (t.co)
A free no-ads app that reads transit cards and maps trips is a small but durable example of useful civic-adjacent software
31.
Quickly whipped up a few CuteDSL kernels for sm120 (TMA+mma.sync). There is quite a gap wrt CuBLAS @ 600W, but su…
Quickly whipped up a few CuteDSL kernels for sm120 (TMA+mma.sync). There is quite a gap wrt CuBLAS @ 600W, but surprisingly, my kernels come up slightly ahead @ 400W. I have a feeling TMA is more energy-efficient than cp.async? (CuBLAS disp
32.
For days, many folks here are citing DeepSWE as the benchmark that restores reality only because it shows GPT 5.5…
For days, many folks here are citing DeepSWE as the benchmark that restores reality only because it shows GPT 5.5 on top. But actually, it almost gets a single entry right: the top one, and all the rest is shuffled.
33.
We will be presenting our CVPR paper When Pretty Isn’t Useful in the afternoon today. (t.co)
We will be presenting our CVPR paper When Pretty Isn’t Useful in the afternoon today. We benchmark modern t2i models and find that newer models make progressively worse synthetic training data and we investigate why! Come visit us at post
34.
Happy Sunday (x.com)
Happy Sunday The king of stealth finally revealed itself @tnforestdept @ntca_india @NatureIn_Focus @NatGeoPhotos @SonyAlpha @atrpollachi
35.
Fintech systems process millions of transactions daily.
Fintech systems process millions of transactions daily. Every payment needs two things to happen: → Save transaction to DB → Notify downstream services (ledger, notifications) Looks simple. But in microservices, this is one of the hardest p
36.
Investors will still invest in stuff like this, yes. We had no real customers or regulatory approval for around 1…
Investors will still invest in stuff like this, yes. We had no real customers or regulatory approval for around 18 months, and still raised ~85m pre revenue and another $30m our first month of revenue. You just have to explain why what you’
37.
(1/3) (x.com)
(1/3) Your daily reminder that lines of code are not a good measure of productivity! Below I've plotted Claude outage frequency (across all products) by quarter over time. @AnthropicAI : Code contribution per person has sure gone up,
38.
Back in March, I tested whether Qwen3.5 4B was as good as GPT-4o for gen chat. The result was that they were ~sim…
Back in March, I tested whether Qwen3.5 4B was as good as GPT-4o for gen chat. The result was that they were ~similar. But lots of models have since come out - Qwen3.6, Gemma4 - how do they stack up? The answer is decisive: they beat GPT-4o
39.
Want to fully audit your LLM harness in your company and stop worrying about key leaks? (x.com)
Want to fully audit your LLM harness in your company and stop worrying about key leaks? I (together with my agent) built a MITM proxy in Go that lets you run @openclaw in Docker with complete traffic audit and key swap. It’s a critical
40.
Apparently our new optimizer SUMO is also better at domain adaptation.
Apparently our new optimizer SUMO is also better at domain adaptation. Small MNIST→SVHN adapter setup: train on MNIST, freeze most of the backbone, then adapt to SVHN with domain adapters + head. SUMO got the best target accuracy and best
41.
Baby bears racing up on tree
Baby bears racing up on tree
42.
What if VR apps did not have to be separate worlds?
What if VR apps did not have to be separate worlds? I have been building Adamas VR around this idea: Small TypeScript projects that can be loaded into the same shared multiplayer VR space. One project could be a point cloud viewer. Anoth
43.
some fireflies I photographed last year
some fireflies I photographed last year
44.
Thorn. Bronze with 925 sterling silver wings.
Thorn. Bronze with 925 sterling silver wings.
45.
Mythos has just recreated Cut the Rope.
Mythos has just recreated Cut the Rope. I'd already shown you a screenshot of the result, but here's a full video demo this time. And that's not all: I asked it to add a complete menu as well as a level editor directly integrated into t
46.
I haven’t flew FPV drones in quite a while. But wow, this sunset at Fort Abercrombie has me falling in love again
I haven’t flew FPV drones in quite a while. But wow, this sunset at Fort Abercrombie has me falling in love again
47.
Built a new component library: Terminal-style agent UI (t.co)
Built a new component library: Terminal-style agent UI ready to use, 3 themes + customizable, supports dark and light mode http:// terminal-kit.com
48.
Hollywood dreams bathed in golden sunlight
Hollywood dreams bathed in golden sunlight
49.
Found a few % parser perf improvement in the parser after telling Claude to follow the ecma spec. (t.co)
Found a few % parser perf improvement in the parser after telling Claude to follow the ecma spec. PR: https:// github.com/oxc-project/ox c/pull/23063 …
50.
Digging deeper into introspection, and there seems to be two types that I don't clearly see differentiated anywhere:
Digging deeper into introspection, and there seems to be two types that I don't clearly see differentiated anywhere: Behavioral introspection: Where a model reasons about its own behavior (predicting what it'll do, or knowing what it knows
51.
the distinguishing feature of the RLM, as expressed many times by Alex et. al. is that its output as code is exec…
the distinguishing feature of the RLM, as expressed many times by Alex et. al. is that its output as code is executed in an environment _which itself has access to model/agent calls_ this is what makes it recursive the dynamic workflow is
52.
This model now in neural-txt library!
This model now in neural-txt library! Quick inference on a bunch of useful tasks, like question generation, answers, extracting bullet points, summarization, knowledge graphs, retrieval etc. Runs very fast, almost too fast at ~250 tok/
53.
two main parallelisms to understand for distributed training
two main parallelisms to understand for distributed training id probably group most of the parallelism to fall under inter-op or intra-op parallelism or both. inter-op parallelism 1. assign different operators to different devices. The se
54.
Worth resurfacing our Constrained Discrete Diffusion (CDD), led by @MichaelCardei and @JacobKChristopher. (t.co)
Worth resurfacing our Constrained Discrete Diffusion (CDD), led by @MichaelCardei and @JacobKChristopher. CDD enforces hard, sequence-level constraints inside discrete diffusion sampling via a differentiable projection: blog: https:// n
55.
oh, right #pitchyagame! ehem...
oh, right #pitchyagame! ehem... PAR52: Balatro + Dorfromantik + Golf = cozy procgen deckbuilder where every card is a stroke Number → distance Suit → skim water, ricochet, chip over Stars → new cards Play 9 holes (or... more? ) #Indi
56.
Two scenarios can look almost identical, yet one is safe and one isn't.
Two scenarios can look almost identical, yet one is safe and one isn't. Can MLLMs tell the difference? Mostly, not well. Our #CVPR2026 paper tackles "contextual safety": not just refusing obviously unsafe inputs, but reading the subtle co
57.
astro-erudite v2 is out! I've rewritten my minimal blogging template from scratch, and it's better than the origi…
astro-erudite v2 is out! I've rewritten my minimal blogging template from scratch, and it's better than the original in every way I know how to measure. It ships with 97% less JavaScript (253kb to 6.5kb) and builds 5x faster (6.1s to 1.2s).
58.
imo the metric isn't impact, it's impact / compute — however GPU-rich you are.
imo the metric isn't impact, it's impact / compute — however GPU-rich you are. learn to do more "offline RL" (e.g. offline analysis like debugging agent traces), where humans are still much better than AI. Before compute stops being subsid
59.
Google Maps lets you look at places (x.com)
Google Maps lets you look at places with world models you can experience them Built smt where you can click anywhere on map, drop in with Lingbot, and walk around. Made with @reactor
60.
Executing this 2 tranche deal without publicity and disclosing to all investors in docs is fine.
Executing this 2 tranche deal without publicity and disclosing to all investors in docs is fine. Publicly announcing "we raised $1B @ $5B from Sequoia" if it isn't true is securities fraud.
61.
SARDI introduces a training-free self-augmenting retrieval framework for dLLMs.
SARDI introduces a training-free self-augmenting retrieval framework for dLLMs. Instead of treating low-confidence tokens as noise to be discarded, it shows that these tokens often contain useful lookahead signals about the final answer.
62.
Opus 4.7 is as good or better than ChemDraw at interpreting NMR spectra on our evals. We're making Claude more he…
Opus 4.7 is as good or better than ChemDraw at interpreting NMR spectra on our evals. We're making Claude more helpful for chemists, starting with routine and time-consuming analytical tasks.
63.
Let's now try reverse-engineering 8 Sleep to work without subscription.
Let's now try reverse-engineering 8 Sleep to work without subscription.
64.
@riverflow_ai (x.com)
@riverflow_ai : We ran SynthID detection on images from Riverflow 2.5. Some carry Google's SynthID watermark. Others carry OpenAI's SynthID signature. Why does a "first-party model" output watermarks from two different providers? Riv
65.
I've released the source code for the CSS animation of likes (t.co)
I've released the source code for the CSS animation of likes https:// github.com/yui540/css-ani mations/tree/main/2026-06-07 …
66.
If you enter the same image hallucination prompt to ChatGPT but in Chinese, it’d generate suggestive photos instead
If you enter the same image hallucination prompt to ChatGPT but in Chinese, it’d generate suggestive photos instead However, it refuses when I prompt it in Japanese or Korean Makes me wonder what kind of bias this comes from / where they
67.
Imagine raising a pristine seed round off a deck about “agentic workflows,” then the buyer forwards one ugly spre…
Imagine raising a pristine seed round off a deck about “agentic workflows,” then the buyer forwards one ugly spreadsheet showing the job already done by two interns and a script. that is the $800k ferrari moment for fake ai startups. chea
68.
there are going to be a few fleeting years before tasteful ASI during which time taste is going to be the most va…
there are going to be a few fleeting years before tasteful ASI during which time taste is going to be the most valuable trait. this is also why taste is going to be hard to engineer through RL, you need some sort of sleep-time compute surfa
69.
HPE server diagrams are equally cool:
HPE server diagrams are equally cool:
70.
Oracle server diagrams are some of the best. They're illustrations, complete with highlights/shadows, 3D-forms wi…
Oracle server diagrams are some of the best. They're illustrations, complete with highlights/shadows, 3D-forms with shading. The lazy thing to do would be to take a picture of it and stick labels over it, usually found in cheap manuals. An
71.
partial erasure + stroke splitting
partial erasure + stroke splitting
72.
Spent three hours reading the inspiring blog. One thing it made me notice: almost every strong flow-map T2I model…
Spent three hours reading the inspiring blog. One thing it made me notice: almost every strong flow-map T2I model ends up adding a GAN loss — CTM, SANA-Sprint, and AYF. In the GAN-finetuning stage it almost feels inverted: the adversarial
73.
Went hunting for why an end-to-end test was so slow. Found a wait loop that looked like it could be prodded by th…
Went hunting for why an end-to-end test was so slow. Found a wait loop that looked like it could be prodded by the test to speed it up. Doing so broke the test! Hours later, I have it: a subtle KVM entry/exit accounting bug in the productio
74.
one way to think about timing your startup launch:
one way to think about timing your startup launch: - if your launch flops, the extra week of polish wouldn't have changed that. might as well learn it, iterate, and try again. - if your launch lands, you'll wish you'd done it sooner. the
75.
credits where it's due
credits where it's due I was working on a feature where we had to topup user credits for a user every month. My instant thought was to use a cron job to make sure it happens in the background when needed
76.
400M tokens spent on DeepSeek V4 Pro trying to make DeepSWE works properly, with the right thinking effort and di…
400M tokens spent on DeepSeek V4 Pro trying to make DeepSWE works properly, with the right thinking effort and direct calls to DeepSeek not OpenRouter 6$ spent so far.
77.
Solar panels are getting the bus bars epoxied down. We will have a bunch of small modifications like this as we c…
Solar panels are getting the bus bars epoxied down. We will have a bunch of small modifications like this as we continue testing. The issue here, as I highlighted a few weeks ago, was due to the fact that it is almost impossible for us to
78.
I have a macOS lume VM set up with a separate Apple account and that's where BlueBubbles runs. Codex on the host … (t.co)
I have a macOS lume VM set up with a separate Apple account and that's where BlueBubbles runs. Codex on the host uses it to iMessage me, and I can iMessage Codex from my phone or watch Here's the gist: https:// gist.github.com/madebydia/7
79.
Exactly one year ago today, (t.co)
Exactly one year ago today, http:// patterncraft.fun came to life 𑣲 Never thought this little thing would take me to so many places and connect me with so many amazing people Forever grateful for all the love, support, and every single
80.
Buy directly from Terminal (Cashfree SDK)
Buy directly from Terminal (Cashfree SDK) Built with Golang & Fastify [Open Source btw] Scan UPI QR & Pay directly from Terminal. We will bring a CLI of Gray Cup so that more programmers can buy our tea and coffee. Inspired from Termin
81.
There’s now also a desktop app which collects all the data from Whoop without subscription.
There’s now also a desktop app which collects all the data from Whoop without subscription.
82.
SAM3 masking by free-text is very useful for drone footage. Masking "sky and water" is important for SfM's accura…
SAM3 masking by free-text is very useful for drone footage. Masking "sky and water" is important for SfM's accuracy and the SfM result was very good.
83.
bullish on 'spec-driven development'
bullish on 'spec-driven development' the frontier coding models have become good enough at long-horizon that i should stop working within codex/claude code. but rather iterate more in obsidian and give the agents large markdown design do
84.
claude down again
claude down again back to gpt 5.5 i've manually achieved a 99.9%+ high availability setup by making my workflow completely agent-agnostic
85.
On a night after the rain, catfish had come to the rice paddies around Lake Biwa to spawn. This was a decisive mo…
On a night after the rain, catfish had come to the rice paddies around Lake Biwa to spawn. This was a decisive moment encountered during a survey of fish species that utilize paddy fields. Such scenes have become almost unseen in recent yea
86.
These are the best spray bottles, it’s an insane price to pay but you get upside down spraying (weighted inlet ho…
These are the best spray bottles, it’s an insane price to pay but you get upside down spraying (weighted inlet hose), a laser-like jet mode, and fuel injector-ass atomization
87.
The Jump targeting an insect showing off the nine colors Indian Pitta.
The Jump targeting an insect showing off the nine colors Indian Pitta. #indiaves #thephotohour #TwitterNatureCommunity #wildlifephotography #birds #birding
88.
i made a little tool that fills in all of your knowledge gaps as you read so you can understand any blog immediately (x.com)
i made a little tool that fills in all of your knowledge gaps as you read so you can understand any blog immediately @portaltext
89.
Beautiful textures in this work from Alex
Beautiful textures in this work from Alex
90.
the real tell of a native new yorker? waits "on line" (t.co)
the real tell of a native new yorker? waits "on line" made a little extension to fix prepositions on streeteasy. go knicks! https:// chromewebstore.google.com/detail/nyc-pre position-fixer/ipmjkopnegbmmoemkcnacbbahnnonpad?utm_campaign=23