1.
The datacenter that built its own gas plant
A datacenter operator bypassed utility delays by building its own gas plant, showing how power infrastructure is now part of software deployment
Backlist — 20 Jun 2026 UTC
Top 90 curated tweets ranked for substance on 20 Jun 2026 UTC.
2.
Data centers lowered electric bills (x.com)
A new analysis says datacenter growth spread fixed grid costs over more kilowatt-hours and modestly pushed national bills down
3.
Why VC struggles to fund hard tech
The article argues that ten-year VC funds are a poor fit for fusion, SMRs, and other long-horizon hard-tech bets
4.
Elliott and SVP buy Braskem debt
Elliott and SVP taking debt exposure in a restructuring signals how distressed buyers position ahead of a deeper turnaround
5.
Twenty raises $100M for a cyber warfare startup
A new $1B valuation for a cyber offensive startup is a reminder that security is now a venture-scale category
6.
Server prices are tripling under AI demand
When core infrastructure costs jump from $40K to $100K, the bottleneck has clearly moved from software to capacity
7.
Training data is becoming a new Wikipedia
As models train on more public traces, internet memory hardens into something closer to a durable encyclopedia
8.
Meta's engineers are becoming data labelers
A major lab shifting engineers toward human feedback work shows how model training is changing inside frontier companies
9.
Databricks and Confluent were built by self-hosting pain
The modern data stack is still a business around how hard Spark and Kafka are to run well yourself
10.
A startup is storing data in bacterial DNA
Data can now be written into DNA and recovered from a tiny lab setup, which is wild and real
11.
At-home DNA/RNA testing is getting closer
Cheap nanopore sequencing points toward consumer diagnostics that know what's wrong before the doctor visit does
12.
Urukka Veedu, a house that grows out of the land
The design shows how material choice and site orientation can make a house feel like a continuation of the landscape
13.
Ed Catmull on Steve Jobs' secret iPhone demo
The story captures how much of Apple's most important work was still hidden even from close collaborators
14.
1973 Istanbul: local TVs cost twice as much
The price gap is a sharp snapshot of how local assembly and import dependence shape consumer markets
15.
Austria's first nuclear plant never opened
A completed reactor was voted down before it ever ran, turning energy policy into a historical cliff edge
16.
Doctors are still fighting population screening
The debate centers on whether false positives and overtreatment make broad imaging worse than doing nothing
17.
MRIs are warping medicine
When a scan becomes expensive enough, the whole clinical system starts optimizing around the machine instead of the patient
18.
How A16Z New Media compounds attention
Distribution, goodwill, and portfolio outcomes compound when a firm becomes a channel for interesting work
19.
A $10M-$30M RFP for AI's real-world impacts (x.com)
The funder is putting serious money behind research on how AI may hit economies, health systems, and science
20.
Tokenized stocks are converging on a stablecoin model
Separating KYC minting from permissionless transfers is becoming the dominant structure for onchain equities
21.
A Morpho strategy with 100% APY and $17M exposure
Full utilization and triple-digit APY are the kind of numbers that usually hide serious fragility
22.
Parcel's file watcher quietly powers the JS stack
A single file-watching component sits underneath a lot of modern dev tooling, which is exactly the kind of invisible infrastructure HN loves
23.
Cloudflare fixed a painful Durable Objects bug
Keeping active WebSockets alive removes a failure mode that matters immediately for streamed agents and long-running tasks
24.
Starting from first principles can expand reasoning and self-delusion
More chain-of-thought can mean more insight, but it also gives models more room to rationalize bad answers
25.
DeepSeek's jobs page quietly points to a hiring push
A job posting can be the earliest public signal of where a fast-moving lab is headed
26.
Z.ai open-sources its RL stack (t.co)
Publishing the full post-training stack makes the gap between research claims and reproducible systems much smaller
27.
The supplier-side gateway problem for AI agents
Agent networks only scale if vendors get a safe, structured way to respond without exposing their internals
28.
Codex should record microphone audio
Capturing spoken intent during tool use would turn narration into a first-class part of agent memory
29.
Under the reflecting pool
Removing a section exposed the physical layers beneath one of Washington's most symbolic public spaces
30.
Silence of the Lambs as a male power study
Clarice is framed as the outsider in nearly every interaction, which is what makes Lecter feel like the only safe conversation partner
31.
New blogpost on using cheap proxy rewards to train models (80+ rl runs) and translating this still into real capa…
New blogpost on using cheap proxy rewards to train models (80+ rl runs) and translating this still into real capability gains. Using some automated environment engineering and what i call "traps" placed around. It's been a fun experiment fo
32.
More money/compute != better models
More money/compute != better models Even GDM has now fallen behind as Anthropic storms fwd with top tier data from all tough prompt it has received as Claude Code dominated professional usage. Meta turned its engineers into engineering data
33.
Agents are motivating so many healthy software habits. Open APIs, documentation (skills), tests (evals), Unix (CL…
Agents are motivating so many healthy software habits. Open APIs, documentation (skills), tests (evals), Unix (CLIs), payment & commerce protocols, even wide 𝙰𝚌𝚌𝚎𝚙𝚝 use (markdown/json/html). The original vision of the WWW coming to l
34.
Excited to wake up to a meeting invite from Noam! (x.com)
Excited to wake up to a meeting invite from Noam! Reminder: OpenAI is also hiring top pre-training researchers in London. Come work with me @SavinovNikolay and the team on architecture, long-context, science of scaling, and more! Conta
35.
essentialai to nvidia is what i heard
essentialai to nvidia is what i heard
36.
kids be rewriting history again
kids be rewriting history again
37.
GLM-5.2 delivers a substantial leap in app development capabilities, which also represent demanding long-horizon …
GLM-5.2 delivers a substantial leap in app development capabilities, which also represent demanding long-horizon tasks. Results: - GLM-5.1: 21/70 - GLM-5.2: 48/70 - Claude Fable 5: 56/70 That's more than a twofold improvement from GLM-5.1
38.
Infrastructure-as-Actors is an unexpected trend I've noticed this year
Infrastructure-as-Actors is an unexpected trend I've noticed this year Including projects like Ramp and OpenClaw Ansible & Terraform are built for static infra that's deployed by a human & doesn't change frequently Agents shifted the nee
39.
GLM 5.2 (max) scores 70.1% on WeirdML, narrowly beating to Genini 3 Pro, from 7 months ago.
GLM 5.2 (max) scores 70.1% on WeirdML, narrowly beating to Genini 3 Pro, from 7 months ago. It uses ~22k output tokens on average, compared to ~12k for the (high) setting. This gives a fairly clear but modest increase (3%) in score, showi
40.
Riverton House by Alexandra Buchanan Architecture
Riverton House by Alexandra Buchanan Architecture Chelmer, Australia
41.
EssentialAI seems to have been acquihired
Editor’s note: imported_from_x_likes
EssentialAI seems to have been acquihired also crazy to find this out from linkedin, i checked and there is 0 information about it (they were valued around ~$1B in August 2025)
42.
i didn't see anyone post about this, but new deepseek domain was registered that routes to a job posting
i didn't see anyone post about this, but new deepseek domain was registered that routes to a job posting bench(dot)deepseek(dot)com
43.
AI-for-chip-design companies are really good at starting “partnerships” and not great at generating recurring rev…
AI-for-chip-design companies are really good at starting “partnerships” and not great at generating recurring revenue, because the EDA tool sales model is fundamentally limiting
44.
Missing book: Glorious Committees of History, on great committees that accomplished great things as committees.
Missing book: Glorious Committees of History, on great committees that accomplished great things as committees.
45.
Ying Sheng co-wrote SGLang, the inference engine now serving Grok at xAI on a hundred thousand GPUs.
Ying Sheng co-wrote SGLang, the inference engine now serving Grok at xAI on a hundred thousand GPUs. She also built FlexGen, which made a 175-billion model run on a single consumer GPU, and helped build Chatbot Arena. Three artifacts the
46.
Goldman Sachs had it, but SoftBank didn't. I felt that early on in my tenure—it was "corporate culture."
Goldman Sachs had it, but SoftBank didn't. I felt that early on in my tenure—it was "corporate culture." That's because SoftBank has Masayoshi Son as its absolute leader. Mr. Son makes the decisions, and the organization follows. On the ot
47.
Your occasional reminder that causal decision theory agents will pay not to know all sorts of information. Eg, p…
Your occasional reminder that causal decision theory agents will pay not to know all sorts of information. Eg, pay not to be told about an Adversarial Offer (Oesterheld and Conitzer).
48.
Intel with another huge hire. Former SK hynix CEO
Intel with another huge hire. Former SK hynix CEO $INTC
49.
More than a year ago (x.com)
More than a year ago @TacoCohen in SPO https:// arxiv.org/abs/2503.05453 already derives a critic-free value/Q parameterization from the policy-reference log-ratio under KL-regularized RL by exactly “deriving value loss through policy ra
50.
Vlad was not a hero,
Vlad was not a hero, Vlad was a Christian killer. He invited thousands of Orthodox Christian Wallachian boyars to an Easter feast and impaled them. Indiscriminately slaughtering women, children, and the elderly, he massacred about 30,00
51.
If you spend too long consuming without producing, you fall into the state of "blah" -- restless yet too tired to…
If you spend too long consuming without producing, you fall into the state of "blah" -- restless yet too tired to do anything. The fix is active production, not more passive consumption.
52.
This is a challenge unique to carve-outs, isn't it?
This is a challenge unique to carve-outs, isn't it? Even for a carve-out by Supporters, who originally knew the business inside out, it still took considerable cost and effort. The transition of systems and security is one thing, but phys
53.
Josh Home Ownership by YIWUJIE Design
Josh Home Ownership by YIWUJIE Design Guangdong Province, China
54.
チャットプラスという会社が先週、上場承認を受けた。
チャットプラスという会社が先週、上場承認を受けた。 従業員は22人。想定時価総額は約49億円。売上10.2億、経常利益率36%。社長は大手建設会社出身の女性だ。
55.
Every dev tool that has historically had a steep learning curve is now usable again with agents.
Every dev tool that has historically had a steep learning curve is now usable again with agents. Especially if 'in distribution'.
56.
I opened an issue in the Redis repository about the 50% memory savings for sorted sets I talked about a few days …
I opened an issue in the Redis repository about the 50% memory savings for sorted sets I talked about a few days ago:
57.
黄金大门对于在后台偷跑的软件有强制关闭的逻辑了。
黄金大门对于在后台偷跑的软件有强制关闭的逻辑了。
58.
Turns out zstd compression can be used as a language model. It can learn language structure, new words, capture s…
Turns out zstd compression can be used as a language model. It can learn language structure, new words, capture semantic similarity of questions and produce the right answer, and write slightly gibberish poetry. Very lightweight, can run o
59.
I still believe declarative IaC is useful here. Actors can generate Alchemy TypeScript and run it in something li…
I still believe declarative IaC is useful here. Actors can generate Alchemy TypeScript and run it in something like a Durable Object, letting it have its own infra without having the LLM perform the various (potentially destructive) steps o
60.
U-Net
U-Net
61.
I'm happy to announce that Binius64 is now, finally, zero-knowledge!
I'm happy to announce that Binius64 is now, finally, zero-knowledge! Until now, Binius64 was a demonstration of the performance binary field systems can deliver. With the ZK functionality, it can be used for privacy applications.
62.
agent wallet hot take:
agent wallet hot take: crypto people keep selling “the agent can pay.” cool. can it get a refund, respect spend policy, prove authorization, and survive the CFO asking why it bought 400 seats of some cursed SaaS at 1:13am?
63.
People in finance do not dress as they do because it's "part of the forgiveness for wealth in finance"
People in finance do not dress as they do because it's "part of the forgiveness for wealth in finance" People who go into finance do not believe wealth is something that requires forgiveness at all They dress as they do because *absolutel
64.
フランチャイズ中心から直営中心に切り替えた理由は、きちんと説明できた方がいい。
フランチャイズ中心から直営中心に切り替えた理由は、きちんと説明できた方がいい。 直営は重い。 でも、収益性・品質管理・オペレーション改善の自由度が高い。 買い手から見ると、 「コントロールできる成長モデル」 は魅力になる。 戦略転換は、理由までセットで価値になる。
65.
Warriors fan wife: The Knicks winning has made everyone in New York nice for a week. It's like they're microdosin…
Warriors fan wife: The Knicks winning has made everyone in New York nice for a week. It's like they're microdosing California.
66.
I think you should at least have the decency to go through the code overview yourself and understand the architec…
I think you should at least have the decency to go through the code overview yourself and understand the architecture of what you are proposing. Don't mind agent made PRs as long as the human can write the description
67.
OP is right, anyone can pull off a bowl with pulsating Xenia. They live in the equivalent of saltwater swamps (th…
OP is right, anyone can pull off a bowl with pulsating Xenia. They live in the equivalent of saltwater swamps (the reason they pulsate is to create local water circulation where there is none) and are almost impossible to kill.
68.
Also been commissioned to make my design into a book cover a family WC-quiz book
Also been commissioned to make my design into a book cover a family WC-quiz book First book cover I’ve made, and hopefully not the last So if you’re in Norway and sees this, you know who designed the cover
69.
Our experience at Daydream exactly. (x.com)
Our experience at Daydream exactly. When you are measured on the % of trace failures and not “overall intelligence”, the whole game is different. We spend a significant effort into developing models to decide when we should pull a huma
70.
one of my favourite past-times is visiting my friend’s houses! i made a simple digital map so you can visit some …
one of my favourite past-times is visiting my friend’s houses! i made a simple digital map so you can visit some of them too :-)
71.
Edition 31 of the Canonical Document of a long-running hardcore history MUD I built using Claude and a contextual…
Edition 31 of the Canonical Document of a long-running hardcore history MUD I built using Claude and a contextual knowledge graph across a few repos Marquis de Canossa has finally received the Order of the Golden Fleece during the Spanish
72.
This paper trains RLVR reasoning models on token-level distributional deviations rather than uniform token update…
This paper trains RLVR reasoning models on token-level distributional deviations rather than uniform token updates, to avoid the entropy collapse that uniform updates cause. RLVR improves reasoning but suffers an optimization instability:
73.
why is the mac spotlight app so horribly useless
why is the mac spotlight app so horribly useless why is the top result for "codex" 5 random pictures and a dictionary results and not the codex app that i use every single day.
74.
Tfw you realize DIY instructions can be abstracted as ASTs and stateful instruction sets
Tfw you realize DIY instructions can be abstracted as ASTs and stateful instruction sets
75.
They tried with golf, did not work
They tried with golf, did not work Now, they moved on to soccer
76.
today i'm launching Flashtype, a markdown editor for Claude & Codex
today i'm launching Flashtype, a markdown editor for Claude & Codex use Claude/Codex to edit markdown, track every change, and see the diffs open source. available now.
77.
Hi Jim! Thanks for the analysis but you're missing a couple of things that I think would help a lot!
Hi Jim! Thanks for the analysis but you're missing a couple of things that I think would help a lot! You're using H100 and H800 hours for estimating your multiplier. H100 hours and H800 hours are not equivalent so doing a raw division does
78.
If GPT 5.6 still requires me to say "no fallbacks please" I think I'll crash out.
If GPT 5.6 still requires me to say "no fallbacks please" I think I'll crash out.
79.
How is a crypto system - with no deep web2 relationships or distribution - going to compete with this?
How is a crypto system - with no deep web2 relationships or distribution - going to compete with this?
80.
add a little joy when removing items.
add a little joy when removing items. .removing-item .container { filter: blur(8px); opacity: 0; transform: scale(0.92); transition: filter 280ms ease-in, opacity 200ms ease-in, transform 280ms ease-in;
81.
I'm a weird Greek guy who instead of studying for his uni exams decided he should watch early speeches of Obama f…
I'm a weird Greek guy who instead of studying for his uni exams decided he should watch early speeches of Obama from 2004-2009 to see what all the hype was about He was kinda.....not good?
82.
He refuses to acknowledge that metas strength is acquisition moreso than it is unique development
He refuses to acknowledge that metas strength is acquisition moreso than it is unique development Also jumping off the cliff with the crazy layoffs in my opinion was the wrong move
83.
A year ago, with a mix of excitement and fear, I sent the largest term sheet of my life to an unstoppable founder…
A year ago, with a mix of excitement and fear, I sent the largest term sheet of my life to an unstoppable founder determined to change history. I knew if I didn’t step up, the company might not survive — the ambition was too bold. It’s wild
84.
Mike's right, and it's the thing that's been bugging me for a couple years now. AI isn't only getting more powerf…
Mike's right, and it's the thing that's been bugging me for a couple years now. AI isn't only getting more powerful, it's getting more concentrated. The models, the compute, the data, the money, all of it pooling into four or five companies
85.
Nippon Life, Japan's strongest big unicorn hunter lol
Nippon Life, Japan's strongest big unicorn hunter lol > The company is expected to gain investment income of 100 billion to 500 billion yen through its initial public offering (IPO)
86.
My best guess is that Anthropic has cracked recursive self improvement, which is why the top talent wants to be t…
My best guess is that Anthropic has cracked recursive self improvement, which is why the top talent wants to be there
87.
Exact workflow/prompts.
Exact workflow/prompts. GPT-5.6 Pro in ChatGPT can't generate images. I had to use GPT-5.5 to generate the images, then GPT-5.6 to build the site with placeholder images, then GPT-5.5 again to fill the images in with real ones. I've teste
88.
In Chinese we have a saying, 画蛇添足. It means adding feet to your drawing of a snake (doing more work to make thing… (x.com)
In Chinese we have a saying, 画蛇添足. It means adding feet to your drawing of a snake (doing more work to make things worse). Someone had to decide to write code to add this validation. cc @zapmail_ai
89.
Takeaways so far:
Takeaways so far: - It LOVES smooth scrolling and scrolljacking techniques. I'm not a fan of how it feels. (Admittedly I did ask for motion and transitions while scrolling, but you can do that in a way that feels good) - It insists on the
90.
OK, I tried GLM-5.2 and this is a good model.
OK, I tried GLM-5.2 and this is a good model. Probably the first model good enough to eschew closed models from your workflow entirely (except if you need vision). I know this won't run on your laptop, but what are the best current vllm/