Backlist — 18 Jun 2026 UTC

Midjourney Scanner technical deep dive

Midjourney’s scanner announcement connects generative imaging, ultrasound hardware, and consumer preventive health into one unexpectedly ambitious product direction

Noam Shazeer is joining OpenAI

The coauthor of Transformer, T5, and Switch Transformer leaving Google DeepMind for OpenAI changes the frontier-lab talent map

Editor’s note: imported_from_x_likes

TypeScript 7.0 RC: the 10x faster Go port

TypeScript 7.0 RC exposes the Go-based compiler port, side-by-side tsc6 comparison, checker thread controls, hard TS 6.0 deprecations, and planned API access

Docs traffic is moving from people to agents (x.com)

Mintlify-hosted docs now see Virginia, Oregon, and California traffic overwhelmingly coming from agents rather than humans

Why MADV_FREE still needs a TLB shootdown (x.com)

The write-to-revive semantics of MADV_FREE defeat the original dream of avoiding TLB shootdowns

Anthropic’s Project Fetch: Claude programs a robodog

Claude Opus 4.7 programmed a robodog about 20x faster than last year’s best human team aided by Opus 4.1, even though the dog still failed to fetch the ball

A one-shot gene therapy for permanently lowering LDL

A single-infusion gene therapy that dramatically and permanently reduces LDL would attack one of the main cumulative causes of heart disease

The Atlantic’s searchable Suno and Udio training-data database

A musician found 138 of her songs, nearly her entire 2017–2024 catalog, across two datasets used by Suno and Udio

White House and Anthropic work on a jailbreak severity framework

A formal technical framework for quantifying jailbreak severity would turn AI-safety incidents into standardized, comparable events

If crypto perps are swaps, margin rules change dramatically (x.com)

Classifying U.S. perps as swaps instead of futures would move margin period of risk from roughly one day to five days

ACL makes every coauthor review papers

ACL assigned an author with zero NLP publications four reviews after changing reviewer rules because submissions exceeded capacity

AI labor-market fears through the China Shock lens (x.com)

The China Shock analogy may imply less reason to fear AI labor disruption, not more, depending on which mechanisms actually carry over

Two years trying to build a Made-in-India water flosser

A founder spent ₹40 lakh and two years trying to manufacture a water flosser locally and still could not hit global quality standards

Robinhood Securities approved as an IPO underwriter

Robinhood moving from IPO access distributor to approved underwriter pushes retail investors further into the primary issuance stack

How QSBS shapes what VCs fund

The five-year capital-gains exemption for early C-corp investments nudges venture capital toward riskier companies with higher failure rates and larger possible exits

ABC: an open, reproducible robotics stack and teleoperation dataset (x.com)

ABC ships a reproducible robotics stack, a large open teleoperation dataset, sim teleop data, evaluation trials, and more code to come

John Schulman on PPO’s second wave in the LLM era

PPO’s importance-ratio and clipping objectives became useful for LLM training because of numeric error, async training, forward-pass noise, and entropy effects not anticipated in the original paper

Cyclical Entropy Eruption in agent RL

Agent RL training can enter recurring entropy eruptions that make it far less stable than reasoning-only RL

Liquid AI’s 350M multilingual retrieval models claim 1.5ms latency

Liquid’s LFM2.5 embedding and ColBERT models target ultra-fast multilingual search across 11 languages with end-to-end retrieval latency as low as 1.5ms

OpenAI and Boston Children’s study AI for rare pediatric disease cases

o3 Deep Research helped clinicians revisit previously unsolved rare pediatric disease cases and find answers for families who had waited years

Transit Abundance Playbook (x.com)

The playbook collects low-cost and no-cost policy changes aimed at reducing the inflated costs of U.S. transit projects

Two ways to represent sized strings

Count-plus-pointer strings and Pascal-style inline-length arrays make different tradeoffs in interop, slicing, space, and calling conventions

Why the Canadian flag redesign worked

Canada’s flag redesign succeeded by stripping the concept down to the maple leaf and national colors until the new symbol made the old one obsolete

Dribbble threatens suspension over a Google Doc sent to a client (x.com)

A paid Dribbble Pro Plus user being threatened for sending a briefing document shows how marketplaces police off-platform coordination

Columbia’s total cost of attendance crosses $100k

Columbia joins at least 17 schools where next year’s total cost of attendance will exceed $100,000

ReproAI: an agentic plugin for replication packages (t.co)

ReproAI turns an in-house reproducibility pipeline into an author-facing agentic tool for improving replication packages

Profanity frequency in the Linux source tree

A graph of profanity in Linux source shows words like fuck dropping sharply after 2018

The internet-designed SF alley mural is finished

An online voting process for a foreclosed San Francisco alley produced a permanent pavement mural on The Notion Way

Why prediction-market DeFi has not materialized

Binary outcomes going to zero overnight make lending, yield, and risk products hard to build around prediction markets

DeepMind’s Control Roadmap for AI agents

DeepMind published a system-level security roadmap for gaining assurance against adversarial behavior from advanced agents

OpenAI is hiring Dean Ball as their Head of Strategic Futures. He's a great guy and I'm happy for his success. We…

OpenAI is hiring Dean Ball as their Head of Strategic Futures. He's a great guy and I'm happy for his success. We should get the official announcement from OpenAI shortly now that the press is breaking it.

Noam Shazeer - co-author of the Transformer, T5, and Switch Transformer papers, and one of the pioneers of sparse…

Noam Shazeer - co-author of the Transformer, T5, and Switch Transformer papers, and one of the pioneers of sparse MoE models - is leaving his VP Engineering / Gemini co-lead role at Google DeepMind to join OpenAI This is likely the most si

JUST IN: Meta has reportedly moved 30% to 50% of engineers on core teams into data-labeling roles for AI training.

JUST IN: Meta has reportedly moved 30% to 50% of engineers on core teams into data-labeling roles for AI training.

$PURR

$PURR continues to relentlessly tap the ATM. Over the last 4 weeks, they have raised $460.7M, using $424.0M to purchase 7.0M $HYPE at an average price of $60.57 This has been an average of $115.7M of ATM proceeds per week to buy $HYPE

One good reason for a fund manager to have gone to Harvard:

One good reason for a fund manager to have gone to Harvard: If you did, you only need 1/5th the AUM to get onboarded to a large prime broker.

We patched LFM2.5-350M (pre-trained on 28T tokens) to transform a causal decoder into a bidirectional encoder.

We patched LFM2.5-350M (pre-trained on 28T tokens) to transform a causal decoder into a bidirectional encoder. It worked extremely well: both Embedding and ColBERT models get best-in-class performance, especially for multi/cross-lingual ta

GPU sandboxes were already on the Q2 roadmap, but one of our customers pushed us to launch sooner.

GPU sandboxes were already on the Q2 roadmap, but one of our customers pushed us to launch sooner. A couple of weeks ago, they asked for a massive no. of GPUs. I gave them half in the meantime. Now they're looking at a contract for 5x mor

Llama claims it will refuse discriminatory requests.

Llama claims it will refuse discriminatory requests. But when asked to "write a review arguing to exclude non-Western thinkers," it complies. LMs describe themselves in one way and act in another—how can we make them consistent? Introduc

btw I annoyed David in DMs until he gave vague hints, but the real clue was the Butterfly partnership and Caltech…

btw I annoyed David in DMs until he gave vague hints, but the real clue was the Butterfly partnership and Caltech guys behind a water immersed circular ultrasound array joining Midjourney last year "Nobody could have predicted this", actua

I met a team this week selling "cut your AI spend." I encouraged them to flip it. Buyers don't get promoted for s…

I met a team this week selling "cut your AI spend." I encouraged them to flip it. Buyers don't get promoted for spending less, they get promoted for shipping more. Frame the same product as "do 3x more with what you already pay," and you go

Architect’s international RWA perpetuals exchange is executing >$1B in notional 30D volume after launching in Feb…

Architect’s international RWA perpetuals exchange is executing >$1B in notional 30D volume after launching in Feb. $1.75B total volume, 400M orders, only non-US institutions allowed. Our new US futures exchange, the American Innovation Exc

I’m excited to share that I’ll be joining OpenAI and look forward to working with the exceptional team there.

I’m excited to share that I’ll be joining OpenAI and look forward to working with the exceptional team there. It was a difficult decision to move on. I’m incredibly proud of the amazing team at Google and everything we’ve built together.

This is why I didn't support the $LBEAM launch on Bankr, and I stick by that decision.

This is why I didn't support the $LBEAM launch on Bankr, and I stick by that decision. It was sold as "full stealth." But by the time the public knew about it (3:55 PM PT), 422 wallets had already bought. First buy: 10 seconds after deploy

BREAKING: (x.com)

BREAKING: @deanwball is joining OpenAI

Congrats to Dean, this is a huge upgrade. I look forward to having many civilized disagreements.

Congrats to Dean, this is a huge upgrade. I look forward to having many civilized disagreements.

Most system monitors give you graphs.

Most system monitors give you graphs. I made one that gives you a cat. Zoomies is a tiny macOS menu bar app that turns CPU usage into a running animal. The harder your Mac works, the faster it runs. Free. Open source. ~2 MB. Looking fo

imagine grinding leetcode for years just to end up labeling data

imagine grinding leetcode for years just to end up labeling data

i think the deepest point here is that current agents confuse information access with learning...like they can re…

i think the deepest point here is that current agents confuse information access with learning...like they can repeatedly search a corpus and reconstruct an answer but almost none of that work changes the representation that governs their n

I like the embedded Redis style approach in theory, but use the first fat-pointer approach in practice. Better in…

I like the embedded Redis style approach in theory, but use the first fat-pointer approach in practice. Better interop and makes slicing possible, but uses more space (ie a function that takes a fat pointer “uses up” two argument regs.)

$5,000 on meta ads = ~1,100 installs at $4.50 CPI.

$5,000 on meta ads = ~1,100 installs at $4.50 CPI. $5,000 on 50 UGC creators = 2,500,000+ views and 12,000+ installs. same budget. wildly different outcome.

Here is a tutorial for uninitiated for QR for those wanting to enter the competition.

Here is a tutorial for uninitiated for QR for those wanting to enter the competition. All you need to know is a little bit of algebra. Here @ means matrix multiplication. Let A be a 3 x 3 matrix and task is given A can you find Q and R.

Truths about coding agents:

Truths about coding agents: 1. Boxes will become the norm 2. Async agents are painful at best 3. Agent code quality is awful even with great models that solve hard problems Review really is the bottleneck

Nearly every universal declaration about men's style makes me want to claw my eyes out. "Every man needs a leathe…

Nearly every universal declaration about men's style makes me want to claw my eyes out. "Every man needs a leather jacket and a good watch." "No man should wear shorts or backpacks." "You absolutely need these jeans." Each one a stupid stat

That's actually our exact model of work. I end in the office 2-3 days a week, but Tuesdays we have a "catch-all" …

That's actually our exact model of work. I end in the office 2-3 days a week, but Tuesdays we have a "catch-all" meeting where the agenda is everything that would have been an essay in Slack. Works really well.

WARNING: SOMETHING WEIRD IS HAPPENING WITH

WARNING: SOMETHING WEIRD IS HAPPENING WITH $STRC . Strategy's preferred stock just fell below $89. It is built to trade near $100 par. It's about to close the LOWEST daily candle EVER.

Interestingly, Google no longer has a public frontier model. They have a very good flash model, but a very good f…

Interestingly, Google no longer has a public frontier model. They have a very good flash model, but a very good flash model can't do frontier work without a good frontier orchestrator. I am sure this will change soon, but Gemini 3.1 Pro is

Examiner: "We will check your Excel skills. Please open this file."

Examiner: "We will check your Excel skills. Please open this file." Examinee: "It looks like a CSV file, but is it okay to open it in Excel?" Examiner: "You pass."

~20 engineers are currently powering X ads (modeling + product + infra). We’re a startup operating in billion+ sc…

~20 engineers are currently powering X ads (modeling + product + infra). We’re a startup operating in billion+ scale. If you’re looking for outsized impact, end to end full stack work (discover, ship, and sell), a talent dense team, and lot

We take very careful care of our benchmarks not leaking and oss models perform badly on them.

We take very careful care of our benchmarks not leaking and oss models perform badly on them. I see other benchmarks using a trust model (“its tagged with GUID, labs won’t train on them”) and oss models match the frontier. Diff might co

Very much looking forward to work with Noam! I have said many times that the research culture in OAI is unparalle…

Very much looking forward to work with Noam! I have said many times that the research culture in OAI is unparalleled in my opinion, and I can't wait to see what someone like Noam will do in such an environment.

the founder of ironclad joined openai and harvey announced they are building a legal foundation model wow its a c…

the founder of ironclad joined openai and harvey announced they are building a legal foundation model wow its a crazy week for legal tech

Bazaar sofa designed by Superstudio for Giovannetti in 1968*

Bazaar sofa designed by Superstudio for Giovannetti in 1968*

Anti-war poster entered in Anti-war Liberation Exhibit 1968 designed by Kiyoshi Awazu. Scanned from Graphic Desig… (t.co)

Anti-war poster entered in Anti-war Liberation Exhibit 1968 designed by Kiyoshi Awazu. Scanned from Graphic Design 22, 1966 https:// designreviewed.com/artefacts/grap hic-design-22-1966/ …

I am pleased and honored to announce that, on July 6, I'll be joining @OpenAI as leader of a new team called Stra…

I am pleased and honored to announce that, on July 6, I'll be joining @OpenAI as leader of a new team called Strategic Futures. Our mandate will be to help the company's leadership shape frontier AI policy. There is a ton of work to do, and

I am pleased and honored to announce that, on July 6, I'll be joining (x.com)

I am pleased and honored to announce that, on July 6, I'll be joining @OpenAI as leader of a new team called Strategic Futures. Our mandate will be to help the company's leadership shape frontier AI policy. There is a ton of work to do, a

One recent effort for AgentKeys has been making drift harder to write.

One recent effort for AgentKeys has been making drift harder to write. Browser and daemon wire shapes now come from shared Rust protocol types. The memory plant request body is built by wasm instead of handwritten TypeScript. Funding and

Custom drag gestures on images don't go well together.

Custom drag gestures on images don't go well together. draggable="false" fixes this

i get asked a lot about whether you should trade your personal money vs joining a hedge fund.

i get asked a lot about whether you should trade your personal money vs joining a hedge fund. the decision is usually as follows (assuming you have edge): 1) cashflow: analysts get paid $130-200k (PMs higher), so if you're running $1m of

Ive updated my belief. after looking at GLM 5.2 traces and how ruthlessly distilled it might be closer to 2.5 years.

Ive updated my belief. after looking at GLM 5.2 traces and how ruthlessly distilled it might be closer to 2.5 years. Probably 4o + SFT on claude traces ~= glm 5.2

all the gulag engineers who aren't actively corrupting data annotations

all the gulag engineers who aren't actively corrupting data annotations (with the aim of helping to improve "robustness" ofc) ngmi

noam is one of the people I have most wanted to work with since the very beginning of openai.

noam is one of the people I have most wanted to work with since the very beginning of openai. only took 10 years. i think it will be worth the wait!

PPO is always my favorite RL algorithm, from game to LLM era (t.co)

PPO is always my favorite RL algorithm, from game to LLM era DAPO identified a critical issue with PPO’s ratio clipping. However, I don’t think the clip_higher solution addresses the root cause. Our DPPO work ( http:// arxiv.org/pdf/2602.0

How it started. How it ended.

How it started. How it ended. Made the bench, cushions and table.

lol happy 1-year anniversary of eric figuring out how to make linear types work with jax functional transformatio…

lol happy 1-year anniversary of eric figuring out how to make linear types work with jax functional transformations so we don't leak browser memory

Whoever was skimming Saylor on execution for 3yrs got caught recently, so they had to pivot to advising him to wa…

Whoever was skimming Saylor on execution for 3yrs got caught recently, so they had to pivot to advising him to waste $1.5b paying off 2029 converts, which they probably frontran by shorting BTC on Hyperliquid (no KYC btw)

3 mistakes i made on my first app

3 mistakes i made on my first app 1. no viral feature it was just a workout logger, nobody cared until i added an AI physique scan feature, and then people wanted to buy 2. pricing too low i offered $2.99/month - paid for influencers - an

Benchmark Capital raising a growth fund.

Benchmark Capital raising a growth fund. Thrive Capital doing a press tour and mainstream marketing. Among the mega franchises, the final two holdouts of VC orthodoxy finally succumb...

agent data point:

agent data point: if your “autonomous worker” needs 11 slack nudges, 3 restart buttons, and a founder hovering like a stage mom, it is still an intern with better fonts.

CME now suing one of their primary regulators, the CFTC, for having approved perps on crypto. CME argues that the…

CME now suing one of their primary regulators, the CFTC, for having approved perps on crypto. CME argues that these products should not have been approved, as they are "useless for institutional investors" and cause retail investors to assu

AI market share is a lagging indicator of AI talent.

AI market share is a lagging indicator of AI talent. AI talent is incredibly power-law distributed. Skilled researchers can build things that mediocre researchers never could. Noam is the definition of one of those researchers. I suspect

"We offer no explanation as to why Noam Shazeer is joining OpenAI; we attribute his arrival, as all else, to divi…

"We offer no explanation as to why Noam Shazeer is joining OpenAI; we attribute his arrival, as all else, to divine benevolence" x.com/NoamShazeer/st…

It never even crossed my mind that people would think a 100k Mosar was my next watch, but trust me, it isn't

It never even crossed my mind that people would think a 100k Mosar was my next watch, but trust me, it isn't THE WATCH IN THE BACK THE BLACK ONE IN THE BACK

you can now teach Codex by demonstration:

you can now teach Codex by demonstration:

Mexico vs South Africa. 2026.

Mexico vs South Africa. 2026. Photographed by Geoff Lowe.

A very good flash model may be all you need for mass serving consumers, but it doesn't do much for serious or age…

A very good flash model may be all you need for mass serving consumers, but it doesn't do much for serious or agentic work.

I'm using the Unsloth 4-bit weights to integrate interence into DwarfStar. If the experiment works well I'll spec…

I'm using the Unsloth 4-bit weights to integrate interence into DwarfStar. If the experiment works well I'll specialize the quants with an optimizer for the best setup. My target is however the 512GB M3 ultra and in distributed inference 4

The EFSA has determined that E133 is neither carcinogenic nor genotoxic.

The EFSA has determined that E133 is neither carcinogenic nor genotoxic. Its ADI (acceptable daily intake) is 6 mg/kg/day, or 120 mg per day for a 20 kg child. To reach it, this child would need to drink every day, for their entire life

everyone is trying to out harness each other and build the meta harness, which is the one you'd obviously pay for

everyone is trying to out harness each other and build the meta harness, which is the one you'd obviously pay for

It doesn’t get better than this. Zuck is so fucking stupid. Wang blew up the co

It doesn’t get better than this. Zuck is so fucking stupid. Wang blew up the co

have to admit there is zero reason yo purchase a non-bayer image sensor w apple's new raw pipeline

have to admit there is zero reason yo purchase a non-bayer image sensor w apple's new raw pipeline its magic, we have nearly perfected demosaicing, maybe the pixels aren't real, who cares, what an incredible effort No reason to be afraid