1.
U.S. export control directive blocks Anthropic’s Fable/Mythos for foreign nationals
Frontier model access was treated as a national-security export-control issue, including for foreign-national employees inside the company
Backlist — 13 Jun 2026 UTC
Balanced the Fable/Mythos export-control shock with hardware, robotics, science, markets, design, security, and urbanism so the page is not just an AI discourse digest.
2.
A Transformer with KV cache implemented in RTL on a Virtex-5 FPGA (x.com)
A full microGPT-style Transformer ran at 56k+ tokens/sec on FPGA fabric with no CPU or GPU
3.
Chronos LTV: Uber’s system for estimating long-term causal impact of delays
Uber built a causal system for estimating how short-term operational delays compound into long-term rider value effects
4.
Inspect your PTX and SASS: Blackwell fuses consecutive max.f32 into FMNMX3
A kernel-tuning note shows how high-level fp32 max operations lower into predicate/select PTX and then fuse into a new Blackwell SASS instruction
5.
Apple rewrote TrueType hinting from C to Swift and made it 13% faster
A memory-safe Swift rewrite of a decades-old font interpreter shipped faster while preserving pixel-perfect output across 27 million glyphs
6.
Omnigent: an open-source meta-harness for combining and controlling coding agents
Databricks released a layer above Claude Code, Codex, Pi, and agent SDKs for composing agents, sharing sessions, and applying control policies
7.
Frontier AI access may look more like licensing chemicals than logging into SaaS
The likely endpoint of capability gating is licensed access for specific uses rather than ad hoc “good versus evil” classifiers
8.
Tokenized SpaceX shares traded at a 40% discount because of lockup mechanics
Synthetic equity wrappers can diverge sharply from underlying shares when conversion rights, lockups, and liquidity differ
9.
An interactive camera exposure simulator
A simulator makes aperture, shutter speed, and ISO legible by letting users adjust them and immediately see the exposure tradeoffs
10.
Bumble bees solve a tool-use puzzle with a ball and fake flower
A Science study found bumble bees can position a ball under a fake flower to reach a reward, challenging assumptions about insect cognition
11.
Awesome-UMI: a directory of Universal Manipulation Interface devices
Human demonstration collection is a bottleneck in humanoid robotics, and UMI-style devices are becoming a reusable tooling ecosystem
12.
SR-71, A-12, and D-21 were built before CAD, STEP, IGES, and NURBS
Some of the most advanced aircraft ever fielded were designed and manufactured before modern geometric modeling tools existed
13.
UNCanny: a Windows AppX/InstallService coercion primitive with an LPE angle
A Windows internals experiment found a coercion primitive around AppX and InstallService with local privilege escalation implications
14.
Custom Metal kernels inside Apple Core AI model graphs speed up sparse MoE decode (t.co)
Dropping custom Metal kernels into .aimodel graphs yielded 2.1–3.6× faster sparse-MoE decode at the same int8 quality
15.
A physical Maxwell-Boltzmann distribution made from beads
A bead-based device turns the statistical distribution of gas molecule speeds into something directly visible
16.
Printed page numbers as browser scrollbars
A browser could use page numbers and link positions as a navigational scrollbar, borrowing affordances from printed matter
17.
Before agents, fix your data
Entity resolution and data fidelity remain hard even when agent scale is cheap, because AI cannot replace looking closely at messy source data
18.
On-policy distillation with positive-pressure tokens
Using only tokens where the teacher assigns higher probability than the student can still minimize an upper bound on on-policy distillation loss
19.
Every $20B-defense-budget country can build cheap attack drones
Cheap group 2/3 one-way drones are becoming domestically producible enough that nationalist defense policy will likely proliferate them widely
20.
India built high-rise electrification for double-stack freight trains
India solved a clearance problem that keeps other major freight systems diesel by building a 7.5-meter electrified rail grid
21.
luwu_mjlab: open-source RL training for quadruped robots (t.co)
An open MuJoCo Warp environment lets quadruped locomotion and get-up policies be trained in simulation and validated on real robots
22.
Hedging locked-up SpaceX gains with a zero-cost collar
Employees with large unrealized gains can offset put protection with call upside to lock in value during a six-month lockup
23.
MiniMax MaxProof crosses the human gold-medal threshold
A test-time proof-search system reportedly scored 35/42 on IMO 2025 and 36/42 on USAMO 2026
24.
Gears of War 2 ads had no focus testing, just trust and taste (x.com)
A major game campaign shipped from team confidence rather than focus tests, arguing for taste as an operational advantage
25.
Per-task economics beat “AI for X” categories
Concrete cost-before, cost-after, error-rate, and escalation-rate metrics reveal more than broad software categories or AI positioning
26.
Farmland preservation as an anti-city policy
Some preservation rules impose large economic costs by blocking the development of land that could support high-productivity cities
27.
Contactism: consciousness as an interface
The theory proposes that minds become conscious only through active interaction, treating consciousness as a boundary phenomenon rather than a stored property
28.
SciAgentArena: 200 real research tasks for science agents
A new benchmark targets messy, multi-step scientific workflows across six domains instead of reducing research ability to math, coding, or Q&A
29.
AXIOS: Anthropic is blocked from releasing Fable outside the US; prohibition includes "foreign persons within the…
AXIOS: Anthropic is blocked from releasing Fable outside the US; prohibition includes "foreign persons within the country"; Howard Lutnick tried to pause the release but Anthropic not convinced
30.
fable5 to be diluted to 3.67% level.
fable5 to be diluted to 3.67% level. negotiations are ongoing whether this will take place inside of anthropic with external oversight, or outside of it
31.
No GPU, no CPU - a full Transformer with KV cache as RTL on a Virtex-5 FPGA. microGPT at ~56k tokens/s, fully ope… (x.com)
No GPU, no CPU - a full Transformer with KV cache as RTL on a Virtex-5 FPGA. microGPT at ~56k tokens/s, fully open-source. Thought you'd appreciate this, @reach_vb
32.
Unless this changes, OpenAI researchers on visas need to plan for the fact they’ll probably lose access to intern…
Unless this changes, OpenAI researchers on visas need to plan for the fact they’ll probably lose access to internal models, and therefore their ability to do their jobs moving forward, sometime in the next couple months. I hope the company
33.
At the start of this project I assumed that to fix misalignment we mainly needed to intervene on the RL stage of …
At the start of this project I assumed that to fix misalignment we mainly needed to intervene on the RL stage of training, and SFT didn't matter much - I was pretty surprised to be wrong! I think these results will plausibly change over t
34.
Fable 5, gone but not forgotten (t.co)
Fable 5, gone but not forgotten I love the northern lights, so I built this aurora simulator. I usually customize and then leave it running on my monitor in the evenings, its quite calming. Give it a try: http:// aurora-sim.vercel.app
35.
contrarian ai take:
contrarian ai take: the safest startup is no longer the one with the prettiest interface. it is the one buried inside a disgusting workflow with permissions, exceptions, refunds, audits, and angry humans. beauty is where incumbents and t
36.
0.01 xA and 0 chances created across these 83 passes per Opta's own stats, btw
0.01 xA and 0 chances created across these 83 passes per Opta's own stats, btw Pass completion is a super naïve metric Nice that he didn't lose possession I guess... but all this really says is that Richards was too conservative -- and bo
37.
Long before 30 Under 30 lost its prestige, it lost credibility amongst listmakers at one of their events. If I re…
Long before 30 Under 30 lost its prestige, it lost credibility amongst listmakers at one of their events. If I recall correctly: drinks got spiked, people started losing their senses, someone comes up to them and pitches their solution TO T
38.
just heard about an AI research lab that wanted funding for long-term research, so instead of raising money they …
just heard about an AI research lab that wanted funding for long-term research, so instead of raising money they apparently turned themselves into an HFT/quant shop and are trying to bankroll the research with trading profits and say they a
39.
Yes, great question!
Yes, great question! Pi runs on the host and uses the sandbox as a sandbox only. Same with OpenAI Agents SDK (not shipped). This is the preferred architecture but most harnesses currently don't support it.
40.
Sideshift is desperately hiring in NYC:
Sideshift is desperately hiring in NYC: Head of Strategy: 130-200k Customer Success: 30k-120k Engineers: 20/hr- 400k Come join us
41.
Just occurred to me that Anthropic employees who are not US persons will not be able to use Fable/Mythos, making …
Just occurred to me that Anthropic employees who are not US persons will not be able to use Fable/Mythos, making this plausibly (and to be clear, accidentally) the first regulation on recursive self-improvement.
42.
confession:
confession: i trust an ugly margin table more than 90% of ai decks. show me: - old cost per task - new cost per task - error rate - who got fired from the workflow - who still needs to approve the weird cases if the economics do not fit
43.
Many such cases, VCs absolutely bamboozled by the ability of mid level quant leavers to generate 7 figure arr in …
Many such cases, VCs absolutely bamboozled by the ability of mid level quant leavers to generate 7 figure arr in a few months to bootstrap their new ventures (and no this kind of pnl is not coming from prediction markets before anyone start
44.
BREAKING: Polymarket trader “Latina” put $800k on Switzerland to win by 2 or more goals today
BREAKING: Polymarket trader “Latina” put $800k on Switzerland to win by 2 or more goals today The payout is $1,387,827.12
45.
“The Admin asked Dario to fix the jailbreak or de-deploy the model. Dario refused.
“The Admin asked Dario to fix the jailbreak or de-deploy the model. Dario refused. — In their blog post, Anthropic defended its decision by saying the jailbreak isn’t serious. “ This is crazy. What are we even doing here?
46.
Gemini 3.1 Pro and Gemini 3 Flash have most qualitative behaviors set by SFT, not RL, contrary to my expectations!
Gemini 3.1 Pro and Gemini 3 Flash have most qualitative behaviors set by SFT, not RL, contrary to my expectations!
47.
No you can still train a mythos class model ground up starting from zero for around 3b all in including compute. …
No you can still train a mythos class model ground up starting from zero for around 3b all in including compute. Any one who quotes a higher figure is talking nonsense
48.
yes. here is an obstacle avoidance drone policy squeezing between walls to reach its target, trained on my macbook
yes. here is an obstacle avoidance drone policy squeezing between walls to reach its target, trained on my macbook
49.
vc hot take:
vc hot take: “proprietary dealflow” often means a founder replied because you had mutuals and a blue check energy drink personality. agents make research cheap. now the costume comes off.
50.
Not sure which founder needs to hear this..
Not sure which founder needs to hear this.. but paid partnerships and boosting “views” on X are now negative signal in all VC group chats. I’m hoping this comes to the inflating ARR trend soon, but that might still take another cycle.
51.
Proximity chat in real life. 500 people in a large room wearing noise cancelling headsets that are connected to t…
Proximity chat in real life. 500 people in a large room wearing noise cancelling headsets that are connected to the people nearest to them
52.
There's misinformation around SpaceX causing retirement accounts to "crash"
There's misinformation around SpaceX causing retirement accounts to "crash" Even if $SPCX was added to the S&P, its weight would be around 2.6%, meaning a 50% drop would only have about a 1.3% impact on the overall index The S&P 500 is
53.
If your entire startup was:
If your entire startup was: “we hillclimb a metric you care about-as-a-service” You’d do pretty well
54.
What if Apple's liquid glass icons were full 3d? Weather edition
What if Apple's liquid glass icons were full 3d? Weather edition
55.
RL training set now available in Use Computer!
RL training set now available in Use Computer!
56.
I've been thinking agent transcripts should be committed along with code changes in a way that makes it easy to c…
I've been thinking agent transcripts should be committed along with code changes in a way that makes it easy to chat with the agent that made a change
57.
Future public models could be nerfed for cyber tasks etc in such a way that you'd have to spend ridiculous amount…
Future public models could be nerfed for cyber tasks etc in such a way that you'd have to spend ridiculous amounts of test time compute to overcome that Smart models will always be able to reason their way to security vulnerabilities but i
58.
Front desk, New Kerylos Hotel.
Front desk, New Kerylos Hotel. Greco-futurism, 2030.
59.
> fable drops
> fable drops > we all feel the AGI for 3 beautiful days > dario runs out of GPUs > dario feels very embarrassed because he never has enough GPUs > nearly 1 GW total of AWS compute for Anthropic coming online later this year > hey bezos can
60.
Just invested in a YC startup at $200M valuation. Pre-demo day.
Just invested in a YC startup at $200M valuation. Pre-demo day. The most expensive seed valuation I’ve ever invested in.
61.
Cow Abducting a UFO by Fer Martinez
Cow Abducting a UFO by Fer Martinez
62.
Socotra Island, Yemen
Socotra Island, Yemen
63.
Waymo somehow costs less than UberX during rush hour (even with Surge Saving)
Waymo somehow costs less than UberX during rush hour (even with Surge Saving) In my experience, Waymo is usually priced similarly to Uber/Lyft’s Comfort Electric
64.
Outputting code is not a problem anymore at all. Most of my time is now spent on getting simpler and dumber code.…
Outputting code is not a problem anymore at all. Most of my time is now spent on getting simpler and dumber code. It's crazy to think about that.
65.
while true, prefill performance is still a huge roadblock
while true, prefill performance is still a huge roadblock 50k tokens of input still unusable on local, that was not the case on gpt4 cloud api even if output tok/s matched the UX lag tax is kind of a big deterrent for the time being
66.
Quantum Metal install used to be ~1 GB. Now it's ~50 MB.
Quantum Metal install used to be ~1 GB. Now it's ~50 MB. Turns out you don't need a desktop GUI to design a quantum chip. Who knew.
67.
U are either on Chinese Open model/Huawei stack or are dependent on USG for model/compute access
U are either on Chinese Open model/Huawei stack or are dependent on USG for model/compute access USG won't allow non American entities to have cyber capabilities that could threaten American national security (tbh they won't even allow p
68.
Fable showed up, did not eat or stop working for seven days, communicated like a maniac in a language of its own …
Fable showed up, did not eat or stop working for seven days, communicated like a maniac in a language of its own abstractions, wrote 50,000 lines of the best code I've ever seen and then just disappeared
69.
Weekend fun project - built a gridworld where the model teaches itself. no step labels - just: try things, keep w… (t.co)
Weekend fun project - built a gridworld where the model teaches itself. no step labels - just: try things, keep what worked, distill it back into your own weights, repeat. the only signal is whether you escaped — and how cleanly. http:// d
70.
We worked 16–18 hour shifts producing the first versions of the Falcon 9 thrusters.
We worked 16–18 hour shifts producing the first versions of the Falcon 9 thrusters. To this day, it is still the hardest manufacturing assignment I have ever been asked to run. We ran the first prototype thrusters in South Bend, IN. I sti
71.
Restricting their customer base
Restricting their customer base Slowing down model releases (by at least 3-6 months) Models that cross a certain intelligence threshold are now a supply chain risk for non US enterprises and foreign subsidiaries of US enterprises Doesn't lo
72.
Agents get expensive when one frontier model does everything: read raw logs, recognize patterns, plan the fix. Th…
Agents get expensive when one frontier model does everything: read raw logs, recognize patterns, plan the fix. The split that works: the big model orchestrates, fine-tuned SLMs work as its tools. Cheap tokens digest raw data into JSON; expe
73.
Introducing Adaline 2.0 - The Agent Self-Improvement Layer
Introducing Adaline 2.0 - The Agent Self-Improvement Layer Adaline turns Traces into Behaviors, Behaviors surface Issues, Issues become auto-generated Evals + Data, Adaline then generates new agent candidates and tests them. You review th
74.
Kimi 2.7 ranked 2nd after Fable 5 and before GPT-5 xhigh
Kimi 2.7 ranked 2nd after Fable 5 and before GPT-5 xhigh We have re-run our ErdosBench smoke test on 14 problems with Kimi 2.7, Qwen 3.7 Max, Grok 4.3 and compared it with the top performers from previous runs. Kimi 2.7 is amazingly good.
75.
it's not lack of compute that's the issze. it's that in Europe, it's unthinkable to pay a guy in his mid 20s $600…
it's not lack of compute that's the issze. it's that in Europe, it's unthinkable to pay a guy in his mid 20s $600k salary and give him resources and freedom to train models without having oversight by a committee of gerontocratic professors
76.
I built a post-run recap design studio in SwiftUI
I built a post-run recap design studio in SwiftUI Live animated map, 2D and 3D rendering, full customization. Here's how everything works
77.
Young people historically had to work for free in order to break into a field. For example this was the case in l…
Young people historically had to work for free in order to break into a field. For example this was the case in law until Cravath decided to make a lawyer assembly line.
78.
Notably, the budget panel was comparable with Claude Fable 5 in performance.
Notably, the budget panel was comparable with Claude Fable 5 in performance. A panel of Gemini 3 Flash, Kimi K2.6, and DeepSeek V4 Pro, fused together, beat solo GPT-5.5 and solo Opus 4.8 outright. And it landed within 1% of Fable 5 while
79.
It is very short of the right kind of smart people: anyone with a solid training experience (let alone at the fro…
It is very short of the right kind of smart people: anyone with a solid training experience (let alone at the frontier) is in the low hundreds.
80.
Surprise, surprise. Training a frontier model requires experience. (x.com)
Surprise, surprise. Training a frontier model requires experience. People from Brain who know all the end-to-end training tricks and recipes since 2020 are 100x more valuable than that “exceptional and energetic” new grad who won a few ha
81.
Not at all crazy.
Not at all crazy. 1. For many (most?) important problems IRL, you're collecting data from the wild and cannot sample completions from the base model. Training RMs has had poor-to-mixed success because of distribution shift. If you're worki
82.
alright hear me out:
alright hear me out: ARR-backed securities. we bundle together hundreds of startups whose recurring revenue comes from other VC-funded startups. then i buy credit default swaps on the bundle, or make banks create them and short the entir
83.
A good time to remind people that in my time doing LLM research I feel like a minority of my colleagues are Ameri…
A good time to remind people that in my time doing LLM research I feel like a minority of my colleagues are American citizens. It would be industry destroying to have to rebuild with segregation for frontier ai research to be legal.
84.
Design Engineering Tip:
Design Engineering Tip: Not every pixel should feel equally movable. Add resistance near important boundaries. It makes interfaces feel physical instead of digital. The effect mainly comes from these things: - dragConstraints → defines t
85.
highly recommend spending a few mins computing a small attention distribution by hand btw
highly recommend spending a few mins computing a small attention distribution by hand btw
86.
"Ay, there's no World Cup atmosphere, nobody will go to the matches."
"Ay, there's no World Cup atmosphere, nobody will go to the matches." A Qatar - Switzerland group stage match:
87.
built a tiny RL rock-stacking experiment (x.com)
built a tiny RL rock-stacking experiment trained on 200k sims via PufferLib ( @jsuarez ) observes rock geometry, mass, friction, velocity, stack contour, support intervals, roll torque, balance, etc
88.
Eh not sure a CB should be judged on xA/chances created. But yes he was mostly recycling possession / keeping thi…
Eh not sure a CB should be judged on xA/chances created. But yes he was mostly recycling possession / keeping things ticking over and not the one breaking lines.
89.
direct hit, no survivors
direct hit, no survivors
90.
chrome plugin also normally broken, there’s many extension conflicts (password managers mainly) that inject shit …
chrome plugin also normally broken, there’s many extension conflicts (password managers mainly) that inject shit across the whole DOM and effectively block the extension. prob need way of disabling them without user consent