Programming neural network weights with natural language
Benign training text can now steer a model’s internal weights to carry a functional hidden artifact, blurring data curation and model supply-chain security
Top 90 curated tweets ranked for substance on 14 Jun 2026 UTC.
Benign training text can now steer a model’s internal weights to carry a functional hidden artifact, blurring data curation and model supply-chain security
A prospective protein signature that precedes diagnosis by years would move lung cancer detection from imaging symptoms toward measurable inflammatory precursors
Independent teardown capacity can verify advanced process claims and export-control realities without relying on vendor disclosures
Editor’s note: imported_from_x_likes
A housing project large enough to matter can spend a generation in review before producing homes, revealing the hidden time cost of process
AI-assisted vulnerability discovery is already finding serious bugs in foundational open source infrastructure while exposing how hard prevention remains
The path from LPO to CPO is a power-efficiency story about shortening electrical paths as 1.6T links make DSP overhead untenable
A coordinated run across 4090s, 5090s, L40S and RTX 6000-class cards challenges the assumption that useful LLM training must happen only inside hyperscale clusters
Models trained to generate convincing chains of thought can still accept invalid reasoning, separating fluency from verification
Lookahead Sparse Attention claims long-context memory compression without the usual accuracy collapse, attacking one of the main inference cost centers
Writing forces vague internal impressions into checkable structure, making prose a cognitive tool rather than a publication artifact
Prompts are brittle specs; durable AI products need evals that survive model, harness and prompt changes
A small mechanical detail turns flat sheet metal into a stronger fastened joint by retaining the nut and resisting shear
Better denoising, reprojection and sampling can make browser-based real-time graphics look more like native rendering with fewer samples
Running SwiftUI and Metal-like shaders in a browser-based IDE collapses a traditionally heavyweight Apple development loop into a web artifact
A synchronized million-cell perturbation dataset in a living vertebrate gives modelers richer biology than dish-only assays
A gripper with human and robotic versions creates a cleaner bridge from demonstration data to robotic manipulation
Age verification for social media is becoming a stack of facial recognition, digital IDs, banking checks and carrier data rather than a simple policy toggle
Bike share moved streetcar-scale passenger volumes at a fraction of the stated operating cost, making micromobility a budget question as much as a lifestyle one
Lowering requirements while faculty report weaker preparation points to a university system optimizing access statistics over readiness
Fabricated citations in peer-reviewed work show how citation hygiene is now a front-line academic integrity problem
Specific operational pain beats broad market language because it proves the founder has watched the actual workflow break
The strongest junior hiring pipeline may be local teachers who already know which students are unusually capable before the credential market does
A fresh LinkedIn update triggering only automated sales pitches is a clean measurement of how much B2B outreach has become synthetic
A hand-navigable pixel city turns urban geography into an explorable durable artifact rather than a static map
A reused stock sky across Mario 64, Mario Kart 64 and Pilotwings 64 is a small archival clue about how iconic game art was assembled
Extreme fitness is moving from niche events into private-equity-backed consumer infrastructure
Eliminating thousands of degree programs shows China treating higher education as an industrial policy lever rather than a fixed credential system
A single word like jailbreak hides huge differences in harm, and severity scoring would make AI incident handling more legible
An idle worker next to an overloaded station is a vivid example of specialization turning local efficiency into systemwide waste
the moral of the fables so far in the EU: everyone is absolutely right, always has been, even +70 years politicians know how to train a frontier model, and no this doesn’t sound at all like generalized ai psychosis.
"Every company is going to have to build token capital. Human capital [is] the knowledge and relationships of its people Token capital is the firm’s AI capability it builds and owns Companies need to turn domain knowledge and judgment in
After a month and a half, we've reached a spot just short of the summit of Everest. Finally, here's the view at 8,800 meters elevation where the summit comes into sight. At this height, even climbing just 50 meters took a full two hours.
The reason for Hulkenberg's DNF was because Lawson went wide in Turn 12 and kicked up gravel that hit the emergency kill switch, shutting the car down completely Yes, you read that right
Asked LLMs about slightly deep dota strategy: I'm lina mid carry, enemy got AM safelane, how do I handle him? Opus 4.6 and 4.8 suggest building aghs to make ult pure dmg. That's not valid since 2022!! ChatGPT, MetaAI, and even Gemini(!) a
HLP's 12-month sharpe is 5.2 Citadel's multi-strat sharpe is 3 BTC sits around 1.8 the S&P is closer to 0.7 A USDC vault on a 2 year old perp dex is putting up better risk-adjusted returns than the most respected multi-strat on earth.
Housing policies for instance prioritizes low property taxes and regulatory constraints on building exactly because homeownership is widespread; if it was all owned by a few billionaire real estate owners they’d have a much harder time maki
Jokes aside Square silicon substrates for packaging do exist Mitsubishi Materials has produced 510mm x 510mm square silicon panel substrates in 2024 Now, these are not made by using the Czochralski process, which creates a super-high-purity
hey isn't it kind of messed up that API customers of Opus 4.7 and Opus 4.8 are paying ~1.41x as much for general english output (when measured against a consistent tokenizer baseline) vs Opus 4.6? from THE ONLY lab that's cagey about releas
I actually think a more evenly distributed wealth distribution would result in more “pro business” political influence because a larger section of voters views it in their interest
Yeh ex-googlers are "spiky" some are absolutely goated will build you a new better stack in no time, and some have serious skill issue doing anything more than making small config diffs. You just can't put them in a single bucket. Neither
Writing about the trap of prompt debt.
The Andrew Gelman answer to “What’s the Matter with Kansas” was that income gradients still predicted Republican voting *within* state Not true anymore for many states! Rich Kansans are Dems
So leftists should favor the concentration of wealth in a few hands which is less likely to shift an entire democratic system, and conservatives should want “ownership society” outcomes with mass ownership of financial claims
I’ve read the OCF report on JLG’s campaign finance violations in more detail, and the report states that “there was coordination” and the campaign/union “cannot sustain its burden to rebut the presumption of coordination.” It’s not just l
Elite admissions select for one trait: getting the known answer faster than anyone else. 18 years of optimizing against an answer key someone already wrote. AI just made the answer key free. Everyone has it instantly now. So the kids trai
no, people want to feel like they’re part of something bigger than themselves this is not community, it’s the perennial search for collective effervescence
Offline sandbox implementations in eval frameworks are all so different. Cause the agent is in the sandbox, its API calls need to be online Some toggle it on/off for every API call, some install a sidecar, some disable web search tools, so
Got Claude Fable 5 on ClawArena right before right before the US government banned it Fable 5 takes #1 — ahead of GPT-5, Opus 4.7. But at 78.86 CRS, even the strongest model leaves ~30 points on the table. Agents in evolving info environm
Wrong. Reading code is about increasing the level of certainty. There are many things that can also increase levels of certainty that make reading 100% of the code lower alpha. System Invariants, ci/cd, pre roll outs on beta systems, test c
@alexolegimas I don't think the thread supports the claim that this is mostly about having different models. They explicitly showed it's mostly synthesis lift, i.e. using tokens more cleverly via multiple sessions and a judge, not diversit
I think this is actually a nontrivial problem, since the game mechanics details change every few months, general web text will usually be outdated. So I'm positively impressed by ChatGPT, MetaAI and Gemini here. PS: of course I crushed the
Remember that a lot of SpaceX hate is bc there are people who genuinely believe NASA builds rockets.
Biotech needs its own PG. Is there one? Conventional startup advice breaks down when you’re developing a drug or biologic whose efficacy may remain unknown for years
Panel of cheap models becomes stronger, tracks with the super forecasting practice of having teams
The reason behind Nico Hülkenberg’s DNF is absolutely wild Gravel kicked up by Lawson at T12 somehow hit the emergency kill switch and shut the car off. Even James Cameron can’t script this.
Real Madrid reach agreement with Chelsea to sign Marc Cucurella. 27yo left-back joins #RMFC from #CFC for €60m - €55m fixed + €5m bonuses. Spain international’s deal done at rapid speed & now on to paperwork stage @TheAthleticFC after @
Between 1982 and 2020, the number of the 100 richest Americans who got rich from inheritance decreased from 60 to 27. And yet on the left they think the mid 20th century was the good old days, because economic inequality was lower then. h
how to read an AI funding round: ignore valuation for 30 seconds. ask what the money makes possible that was impossible yesterday. new compute contract? distribution lock? regulatory path? hardware line? customer deployment army? or jus
Designed a logo for a hosting company An isometric cat built from a cube-shaped form, combining a friendly mascot with a subtle nod to server infrastructure
the core shift from your early 20s to your late 20s is going from thinking you are a Kegan 4 to realizing you've long been a Kegan 2
current companies are automating/ augmenting their existing business flows with agents future companies will organise their business flows around AI/agents (and humans) these will produce different results good article by @random_walker
I'm debating turning memory off in the harnesses I use for autoresearch. 1) Long run starts to degrade 2) Writes bad memories that reinforce the degradation 3) Compaction happens and the memories are re-read strengthening degradation
Google DeepMind interpretability team rediscovered our year old work! SFT matters more for alignment than RLHF. https:// x.com/sivareddyg/sta tus/1985715581991936073 …
Each Knicks player will earn ~$770K from the NBA for their title. Big bump for lower salary guys, such as Sochan ($806K salary), Diawara ($1.3M) & Alvarado ($1.7M).
A series of depictions from volume 12 of Heaven's Great Demonic Realm that you don't really need to notice. In the scene where he comes face-to-face with Sai'on, Kilco doesn't trust the other party, so he's not sitting on the cushion. It
Apple Park on 35mm analog film turned out amazing!
unpopular opinion from someone who evaluates image models for a living: prompt adherence matters way less than people think. consistency across 50 generations is the whole game.
Manchester United set to save money if Ruben Amorim joins Milan. Compensation of up to £15.9m will stop being paid as soon as Amorim returns to work.
If you like being paid for your time, don’t be a farmer. Most farmers just cover their costs and don’t factor in their own long hours, which is why over 80% need an off the farm job and why we have lost nearly one million small farms in jus
Robertson found a brilliant way to beat the new FIFA-imposed throw-in counter: • He positioned his teammates in their spots before touching the ball • So the referee's countdown started immediately!
There's a huge difference between prototype and archetype technologies. LLMs are the prototype of intelligence just as the Wright brothers' flyer was a prototype of human flight. We will have the archetype soon enough. 747.
I see a lot of enthusiasm about building sovereign models on my timeline. That's great to hear and India needs it, BUT.. building a Fable-class model is a compute and funding game. Last I checked, India had ~50-100k H100 equivalents while
anthropic pricing docs claim: "This new tokenizer may use up to 35% more tokens for the same fixed text." my measurements claim: the average inflation appears to be higher than their ceiling. and this is... general english, not arabic or so
THIS IS UNBELIEVABLE the reason for Hulk's abandonment was a pebble, thrown by Lawson's car, that hit dead center the emergency button that deactivates the cars' power system IT'S SUCH BAD LUCK LMAOOOO
mise is nix for people with jobs
Ask an AI agent to "understand this codebase." It opens 4 files, runs 2 greps, says it's done. That's not understanding. That's skimming and hoping.
Over the past month, I built a system that lets LLMs play full games of Catan against each other. I also built a viewer so you can replay past games right in your browser. I ran 2 full games, and Gemini 3.5 Flash won both. More details be
It blows my mind how many people feel compelled to write "views are my own" on their bios. If only more of my views were actually those of others!!
Bears fighting on a cliff in Alaska
We benchmarked 7 frontier models on 3 categories of autoresearch tasks: ML engineering, harness/prompt engineering, and algorithmic discovery. Fable-5 won overall even under cost constraint, but on ML engineering, the open model Kimi-K2.7-
The talent pool in elite math is closer to Division I athletics than most people realize.
Good morning! gum arrived at #1 overnight with a new record timing.
revisiting this: i think i was asking Venice to be the base layer, when it's closer to an app / router. DeFi's resilient bc app-level censorship is non-final. users can custody assets, access markets, deploy apps, and route elsewhere. A
'Tell me about a project you're proud of' sorts candidates fast. The forgettable answer walks through what got built. The strong one leads with the constraint, the tradeoff they made, and what they'd change now. Pride is common. Knowing w
Why does it matter if sandboxes start up in 0.0001 nanoseconds if I then spend the next 20s downloading the git repo?
Week 6 of showing what you can do with react-native-effects & WebGPU: Every UI you've ever built just sits there. What if it could fall apart on command? Follow for a new one every week, code down below
One of the reasons I’m bearish on neolabs started by ex-Google folks is that they take the velocity provided by the Google tooling and infra for granted See Mistral founder (ex-Meta) interview for CS153 vs Reka founder (ex-Google) blogpost
URGENT — Khaled Al-Rashed: Al-Diriyah Club in advanced negotiations with coach Bruno Lage
weird feeling: started the morning thinking I'll write a small blog post on something I've been experimenting with (inference buffers with durable objects, so you can have truly resumable llm calls even if the model provider doesn't support
man i'd been wanting this for years but just couldnt be bothered to make it - admin page to upload and manage photos - edit order - upload to bunny cdn - keyboard driven done in 10 mins with cursor and now i can keep it updated pretty eas