1.
Early-career biomedical scientists after U.S. funding cuts
Funding cuts made young biomedical scientists far less likely to remain in academia, stay in the United States, or feel satisfied with pursuing a science PhD
Backlist — 15 Jun 2026 UTC
Balanced toward durable technical, policy, markets, infrastructure, and culture items while limiting the Fable/agent discourse to the strongest substantive entries
2.
A £13k frontier-AI scan of UK government repos found 407 vulnerabilities
The UK government used AISI and NCSC tooling to scan public code repositories and found hundreds of issues, including critical weaknesses exposing services to auth bypass and data exposure
3.
Billing is a time-varying relationship between customers, plans, and entitlements
Pricing changes are hard because billing state evolves across customers, plans, and entitlements, not just customer rows in a database
4.
A 2009 PHP app served 12k requests/minute on 384MB of RAM
A small bare-metal PHP deployment outperforming a memory-hungry modern rewrite is a sharp reminder that software progress often hides large efficiency regressions
5.
The largest irrigated crop in America is lawn
Roughly 40 million acres of U.S. land are irrigated grass that consumes water without producing food
6.
Gary Gensler’s case against federal sports prediction markets
The amicus brief argues that commodity-derivatives law was not designed to federalize sports betting under prediction-market venues
7.
The Pentagon’s 1993 ‘Last Supper’ and defense consolidation
A Pentagon dinner warning defense CEOs to consolidate helps explain the modern structure of U.S. defense primes
8.
Geothermal turbines do not have the gas-turbine supply bottleneck
Geothermal turbines run at much lower temperatures than gas turbines and can be made with conventional steels from a far broader foundry base
9.
University endowments’ SpaceX exposure
UNC and Washington University reportedly have about 10% of their endowments in SpaceX, with several other major universities holding significant stakes
10.
InP substrate exports as an AI hardware choke point
Beijing’s approvals for InP substrate exports show how a narrow upstream materials dependency can become strategic leverage
11.
Samsung Foundry is reportedly developing Neuralink’s next chip on 4nm
A next-generation Neuralink chip moving through Samsung’s 4nm process would connect advanced foundry capacity directly to implantable neurotechnology
12.
A tongue-controlled trackpad for people with paralysis
Tomás Vega’s palate-mounted wireless device lets people with paralysis control phones, tablets, and computers using only their tongue
13.
ReplaySSM: cache inputs, rebuild state on the fly
As hybrid SSM-transformer models become common, avoiding per-step SSM state writes could make decode substantially faster without changing outputs
14.
Benchmark manipulation allegations on RoboArena
RoboArena’s maintainers say they found evidence of benchmark manipulation and changed procedures to preserve fair robotics evaluations
15.
Robotics RL is often imitation learning in disguise
Reward shaping, curricula, initialization, and environment design can smuggle human demonstrations into reinforcement-learning systems indirectly
16.
Deriving quaternions from dot and outer products
A geometric product defined as dot plus outer product naturally yields quaternions and Clifford algebras from vector multiplication
17.
Megaprop: preconditioned optimization across GPUs
Megaprop extends Megatron and TransformerEngine with distributed support for Muon, FOOF, KFAC, Newton-Muon, and MuP across width and depth
18.
KV-cache compression: 74x smaller, 99% of tokens evicted
A compression prototype looked spectacular until long-context quality tests revealed it had simply deleted almost all tokens that future tasks might need
19.
FastContext: a 4B codebase explorer for coding agents
Microsoft’s repo-exploration model offloads code search from the main coding agent, cutting tokens by up to 60% while improving SWE-bench scores
20.
Invisible yellow tracking dots in color laser printers
Color laser printers have long embedded tiny yellow dot patterns that can encode identifying information on every printed page
21.
Mister Rogers’ 22-year collaboration with a child psychologist
Fred Rogers shaped scripts, props, songs, and pacing around children’s nervous systems, showing how deeply researched calm television can be
22.
How repeated video exposure changes what viewers can infer
Modern viewers understand cuts and implied continuity because decades of watching video have trained faster visual inference
23.
Independent craftsmanship as corporate jobs become funding
Elite talent is increasingly treating salaried work as non-dilutive capital for building personal digital estates outside the firm
24.
Intercom’s reboot into Fin and Salesforce’s $3.6B acquisition
Intercom went from stalled growth to an AI product rebuild, rebrand, and multibillion-dollar exit by disrupting its own business
25.
Laser weeding broccoli at the edge (x.com)
Real-time AI can identify weeds in broccoli fields and destroy them mechanically without herbicides
26.
HoloPathTracer: path tracing for phase holograms (t.co)
A wave-optics renderer can encode full 3D visual cues into phase holograms while remaining computationally efficient
27.
Remotion adds interactive canvas editing that writes back to code
Video compositions can now be manipulated directly on the canvas while preserving code as the source of truth
28.
South Korea’s skilled-worker visa for low fertility and aging
South Korea is trying to attract more than 1,000 skilled tech professionals with flexible work rules and a fast path to permanent residency
29.
Ethiopia’s Bishoftu airport plan: 110M passengers per year
The planned airport near Addis Ababa would handle 4.4 times the traffic of Ethiopia’s current hub and rank among the world’s largest aviation projects
30.
Learning should not be sold as pleasure
Students misled into expecting learning to feel pleasant can mistake necessary cognitive effort for failure instead of progress
31.
is the writing on the wall?
is the writing on the wall?
32.
Magpie completely relentless on a cross fox here in Kodiak
Magpie completely relentless on a cross fox here in Kodiak
33.
ai hot take:
ai hot take: the impressive demo is no longer impressive. watched a founder show 9 minutes of agent logs, 14 failed retries, one escalation rule, and a before/after invoice. that was the company. the shiny chat window was just the lobby
34.
Megaprop's PSGD implementation calculates preconditioning matrices along with the gradient, collecting and commun…
Megaprop's PSGD implementation calculates preconditioning matrices along with the gradient, collecting and communicating X.T @ X and dY.T @ dY at the same time we do the gradient on the weights: dY.T @ X, and has first-class support for dia
35.
agent story i keep seeing:
agent story i keep seeing: 1. customer has a 6-step human queue 2. founder automates step 3 3. queue still sucks 4. founder automates handoffs around step 3 5. suddenly the whole department looks overstaffed agents win in the joints betwe
36.
Another important thing: Chinese models are not strong because they distill US models. Distillation of models via…
Another important thing: Chinese models are not strong because they distill US models. Distillation of models via API is *impossible*. If somebody tells you the contrary, they don't understand machine learning:
37.
I made a thing that sprouts an RJ-45 and Type-C power on the perforated board wall.
I made a thing that sprouts an RJ-45 and Type-C power on the perforated board wall. It connects to the SW or Type-C AC adapter through the hidden slit between the board and the aluminum frame, so the cable isn't visible. It's a mechanism w
38.
jobs left:
jobs left: 1. tell agents exactly what to do 2. own a budget line 3. sell access to rich people 4. be so hot people forgive the business model middle management status theater got rugged.
39.
A new Puffin!
A new Puffin! Taken this weekend during a boat cruise around Skomer Island. It was very tough to take photos from a bobbing and swaying boat, but it was better than nothing as I couldn't get on the Island itself as tickets were all sold o
40.
Why does GPT write 5x more code than Claude?
Why does GPT write 5x more code than Claude? As its last act, I had Fable analyze WeirdML data, and the short answer (link to full analysis in reply): "The gap is real code, not comments. Recent GPT models (since GPT-5) build portfolio
41.
This result is because OpenRouter evaluated fusion on web research tasks with tricks and traps based on what agen…
This result is because OpenRouter evaluated fusion on web research tasks with tricks and traps based on what agents find first. FutureSearch saw this in mid-2025 with multi-agents on Deep Research Bench. This works without fusion, just run
42.
Daniel Bitton reveals he ran an experiment: $5,000 on clipping got 64 million views, a $5,000 billboard got 62 scans
Daniel Bitton reveals he ran an experiment: $5,000 on clipping got 64 million views, a $5,000 billboard got 62 scans “We spent 5,000 on the most popular billboard in Toronto, 24,000 people walked by it in a day, it got 62 scans to the QR c
43.
Chinese model companies are emailing me offering $50,000+ in free inference (x.com)
Chinese model companies are emailing me offering $50,000+ in free inference One even said: “Our engineering team can submit a PR to your repo with the provider integration” @simdotai supports thousands of teams building and deploying A
44.
This isn't very true. A big part of the problem is that the labs use the term distillation, which is a general po…
This isn't very true. A big part of the problem is that the labs use the term distillation, which is a general post-training technique, in lieu of a specific issue of jailbreaking the API. (1) There is a second debate of *how* impactful d
45.
counterargument to the no pretraining purist route is that silver’s framing assumes the reward fully specifies th…
counterargument to the no pretraining purist route is that silver’s framing assumes the reward fully specifies the task. in any domain where it doesn’t, and that’s most of the economy, the research question is “what is the minimum human dat
46.
BREAKING: Someone named “fishalive” put $400k on Spain NOT to win vs Cabo Verde at 9% odds...
BREAKING: Someone named “fishalive” put $400k on Spain NOT to win vs Cabo Verde at 9% odds... This trade just cashed out $4,702,769.23 on Polymarket
47.
Yesterday I encountered the biggest AI alignment issue I’ve ever experienced, when I asked Claude to double check…
Yesterday I encountered the biggest AI alignment issue I’ve ever experienced, when I asked Claude to double check my post. It changed this quoted subpost into a version that was much more favourable to Anthropic, while keeping most of othe
48.
Nest and Cascades by Zaha Hadid Architects
Nest and Cascades by Zaha Hadid Architects Tirana, Albania
49.
I made a piggy bank out of a gutter.
I made a piggy bank out of a gutter. Putting money in it makes me feel like I've lost out.
50.
vc hot take:
vc hot take: half of “proprietary dealflow” is just rich guy group chat cosplay with a calendly link. actual edge is saying yes before the room has social permission.
51.
Another family of 4th-degree Thue inequalities
Another family of 4th-degree Thue inequalities | x⁴ + 4ax³y + 6ax²y² + 4a²xy³ + a²y⁴ | ≦ a² also seems to have already been solved for a≧205, but when I checked the primitive solutions for 2≦a≦204, the only solutions were (x, y) = (±1
52.
if we can extend this fixed point method to (p1(y) - p2(y)x), then x_n converges to p1(y)/p2(y), i.e. we can comp…
if we can extend this fixed point method to (p1(y) - p2(y)x), then x_n converges to p1(y)/p2(y), i.e. we can compute any rational function of spectrum, instead of just fractional powers. And rational functions can fit many functions very we
53.
They also have unrestricted access to Blackwell
They also have unrestricted access to Blackwell
54.
DwarfStar now supports SSD streaming in the DGX Spark and Strix Halo, not just in Metal. You can run the Q4 quant…
DwarfStar now supports SSD streaming in the DGX Spark and Strix Halo, not just in Metal. You can run the Q4 quants at decent speed, and even DeepSeek v4 PRO at low speed, or you can run Q2 Flash if you have less than 128GB.
55.
Update 4: HMC-000 BLDC motor controller
Update 4: HMC-000 BLDC motor controller Added SEQ mode: the controller orders the motor to follow a sequence of deg's Accuracy seems incredibly good in these first tests especially with a temp housing. A combination of my own firmware gre
56.
FLUX.2 [klein] 4B running on-device on Mac via Apple's Core AI (t.co)
FLUX.2 [klein] 4B running on-device on Mac via Apple's Core AI Converted straight from Apple's official recipe — it runs unmodified on the stock CoreAIDiffusionPipeline. . 1024×1024 in ~17s · 4 steps. Model: https:// huggingface.co/mlboy
57.
A month ago, I bought a Tencent Cloud annual host, and got a full refund on the same day. Today, by sheer acciden…
A month ago, I bought a Tencent Cloud annual host, and got a full refund on the same day. Today, by sheer accident, I mistyped the SSH address and somehow logged back in... Is my old employer really this slapdash?!
58.
welcome to the team : ) could not be more excited to have you in DeepMind!!
welcome to the team : ) could not be more excited to have you in DeepMind!!
59.
After a rewarding journey at Ai2 (x.com)
After a rewarding journey at Ai2 @allen_ai , I've recently joined Thinking Machines Lab @thinkymachines . Thank you to all my mentors, colleagues, and friends at Ai2 for shaping how I think about AI research and for giving me the opportun
60.
EXCL: Newcastle United head of recruitment Steve Nickson leaving club. Englishman has informed #NUFC of decision … (x.com)
EXCL: Newcastle United head of recruitment Steve Nickson leaving club. Englishman has informed #NUFC of decision - timings still tbc but all parties comfortable with situation. Move to Championship on agenda after 15yrs at St James’s Park
61.
I love the World Cup, but it’s always more fun when you know a little background, a little drama, a little gossip…
I love the World Cup, but it’s always more fun when you know a little background, a little drama, a little gossip around each match. So I made a little web app that gives you context for every game. Meet Gossip Goals. XOXO
62.
Today is very exciting day for us.
Today is very exciting day for us. I'm so happy and proud for what this means for every single one of our team who worked so hard, and I'm so incredibly thankful for everyone who supported us in any and every way along this journey, and
63.
It’s important to know that the social media ban for under 16s is not a ban for under 16s.
It’s important to know that the social media ban for under 16s is not a ban for under 16s. It is a ban on *selected* social media for EVERYONE. Until you identify yourself.
64.
new: nvidia faces more chip rivals every year yet it's ~gaining~ market share on all of 'em in ai inference mkt.
new: nvidia faces more chip rivals every year yet it's ~gaining~ market share on all of 'em in ai inference mkt.
65.
This was a fascinating project - turns out that LLMs inherit a lot of traits from LLMs they're distilled from, in…
This was a fascinating project - turns out that LLMs inherit a lot of traits from LLMs they're distilled from, including in subtle ways without clear semantic meaning. This has pretty interesting implications - safety problems in a model in
66.
i don't get the "why would you buy at ATH" meme, long term bullish stocks should be at ATH pretty often
i don't get the "why would you buy at ATH" meme, long term bullish stocks should be at ATH pretty often
67.
when you don't randomize the episode length:
when you don't randomize the episode length:
68.
deepseek v1 -> v3 (no details in v4 about this) and k2 don't use muP and instead use naive N(0,0.006) initializat…
deepseek v1 -> v3 (no details in v4 about this) and k2 don't use muP and instead use naive N(0,0.006) initialization. so how do they do hyperparam selection? they basically fit scaling laws to get optimal batch size and learning rate. ther
69.
BREAKING: Someone named “flickraw” just put $1.5M on Belgium to WIN their match vs Egypt today
BREAKING: Someone named “flickraw” just put $1.5M on Belgium to WIN their match vs Egypt today This pays out $2,403,050.26 on Polymarket
70.
After 4 yrs at Google (Firebase & Gemini CLI), today I'm ready to announce what's next (x.com)
After 4 yrs at Google (Firebase & Gemini CLI), today I'm ready to announce what's next Member of Technical Staff, Google AI Studio, reporting to @OfficialLoganK Never thought I'd end up here friends... but work hard, be kind & find peo
71.
There is a large scale scam campaign impersonating the Morpho recruiting team. These emails falsely invite candid…
There is a large scale scam campaign impersonating the Morpho recruiting team. These emails falsely invite candidates to interviews and ask them to install third-party meeting software. They are not from us. Morpho will never ask candidat
72.
“Solving hardware” in any meaningful sense looks less like LLM schematic generators and more like building severa…
“Solving hardware” in any meaningful sense looks less like LLM schematic generators and more like building several hundred PCB factories in Newark NJ
73.
We worked with (x.com)
We worked with @lmsysorg and http:// z-lab.ai to - integrate DFlash spec into @sgl_project - make it faster with overlap - train a DFlash drafter for @Alibaba_Qwen 397B-A17B The result: up to 4.3x greater throughput over baseline an
74.
this is the exact same lr schedule as deepseek v2 in terms of when the lr changes
this is the exact same lr schedule as deepseek v2 in terms of when the lr changes the difference is that the second phase does seem to be a slight decay instead of constant, and of course cooldown is also a decay instead of constant
75.
Spacex now 180 is trading folks , I gave at 150.
Spacex now 180 is trading folks , I gave at 150.
76.
BREAKING: Claude Fable 5 is likely distilled from Mistral Le Gros Chaton. Prompting in French makes it introduce …
BREAKING: Claude Fable 5 is likely distilled from Mistral Le Gros Chaton. Prompting in French makes it introduce itself as such with a strong "cat persona". We've seen this before in Chinese, where Claude introduces itself as Deepseek. We d
77.
A sparse autoencoder aims to learn a dictionary of interpretable features from a model's activations — but a lot …
A sparse autoencoder aims to learn a dictionary of interpretable features from a model's activations — but a lot can come out "dead," never firing once. On some models this is rare; on others, >70% die even with fixes like AuxK. We went dow
78.
Flock is not the reason for more data centers.
Flock is not the reason for more data centers. Flock runs all its plate reads using <0.1% of one data center. There are 4000+ in the US. For every fun AI video, Flock has alerted on ~95 sex offenders, ~85 car thieves, ~85 wanted people, a
79.
Spoke w/ a very senior guy at
Spoke w/ a very senior guy at $META who told me they had this leader board for Token spend. Since they have an "unlimited budget" for AI spend, they track the daily token budgets etc. The top guy at $META had spent ~$90K in a single d
80.
People often have this idea of Chinese models being N months behind US models. This mental model is not helpful t…
People often have this idea of Chinese models being N months behind US models. This mental model is not helpful to predict the future. The lag is due to compute deficit, so the playfield is that. It's not by chance that OpenAI and Anthropic
81.
Congrats (x.com)
Congrats @destraynor , @eoghan and the entire Fin / Intercom team! Very well deserved for impressive execution.
82.
Strategy has acquired 1,587 BTC for $100 million to increase our
Strategy has acquired 1,587 BTC for $100 million to increase our $BTC Reserve to ₿846,842. We have also increased our USD Reserve by $100 million to $1.1 billion. $MSTR $STRC
83.
Don't forget the mad lass that was using 0.4 weight decay in 2021
Don't forget the mad lass that was using 0.4 weight decay in 2021 𝘇𝗲𝗿𝗼 𝗽𝗼𝗶𝗻𝘁 𝗳𝘂𝗰𝗸𝗶𝗻𝗴 𝗳𝗼𝘂𝗿 And nobody batted an eye?!
84.
I made a book jacket for an assignment (I sewed it).
I made a book jacket for an assignment (I sewed it).
85.
Google has started mixing ads right into the middle of image search results. It decreases search quality dramatic…
Google has started mixing ads right into the middle of image search results. It decreases search quality dramatically. Now when you search for images of a watch, you get watches that aren't even from the same manufacturer. I've actually swi
86.
kinda crazy this is the first time in months anthropic employees will no longer have access to mythos internally
kinda crazy this is the first time in months anthropic employees will no longer have access to mythos internally
87.
probably the best blog i have read for some time
probably the best blog i have read for some time viewing SFT, RL, and OPD as different ways of reshaping a model's distribution makes their tradeoffs super intuitive. - SFT pulls toward a fixed external target - RL moves along the reward
88.
“You obviously don’t know how git is implemented”
“You obviously don’t know how git is implemented” Yes, I haven’t read much of git’s code. And that’s still true - downloading a big repo should not be much slower than downloading a big file.
89.
Codex hit Figma's MCP limit and then just opened Figma in a browser tab to keep working
Codex hit Figma's MCP limit and then just opened Figma in a browser tab to keep working
90.
Iturbide Studio by Taller | Mauricio Rocha + Gabriela Carrillo
Iturbide Studio by Taller | Mauricio Rocha + Gabriela Carrillo Mexico City, Mexico