Evgenii

@SethiLiam AP x Swatch is proof that luxury doesn't have to be gatekept. love to see it

Audemars Piguet x Swatch might be the smartest watch collab ever AP gets a younger audience Swatch gets unlimited hype Resellers are about to make a fortune $40k Royal Oak aesthetic for ~$400 was always going to break the internet Not gonna lie… I’m probably trying to grab a few on release day 😭

@anishmoonka the brain running on different firmware just makes the connection stronger.

Scientists ran a debate experiment on 30 four-person groups of strangers, scanning everyone’s brain in real time. The pairs with similar autistic traits clicked together. The brain region powering that connection was completely different from the one the other pairs used. The work came out of Tianjin Normal University and appeared in Biological Psychiatry in September 2025. Researchers used small light-based scalp sensors to capture brain activity as it happened. They put together 30 four-person groups, each with two students scoring high on a standard autism survey and two scoring low. The groups listened to an audio story, then debated which fictional characters to rescue from a deserted island under turn-taking rules. Afterwards, every participant privately rated who they wanted to be friends with. The mutual attraction only showed up when opinions matched during the survival debate. When two students agreed on who to rescue, the one with high autistic traits wanted to be friends with the other high-trait student. The same happened for low-trait students. Matching on personality, like both being extraverts, had no effect. The brain scans revealed the mechanism. Low-trait pairs synced up in a region called the right temporoparietal junction, the social autopilot that reads the room without conscious effort. High-trait pairs lit up elsewhere. Their brains synchronized in the right dorsolateral prefrontal cortex, the focused-thinking region you use to solve a math problem. Both kinds of pairs reached the same level of connection, just through different neural hardware. This finding mirrors a 2020 study from Edinburgh, where researchers passed a story person-to-person through a chain of eight people. Autistic-only chains kept about 13 details out of 30. Non-autistic-only chains kept about 12. Mixed chains, where the two types alternated, dropped to roughly 9 details, a 31% decline. The latest CDC numbers from April 2025 put autism at 1 in 31 US eight-year-olds, up from 1 in 36 just two years earlier. For decades, the standard line in research and clinical practice called this a social skills deficit. School and clinical programs were built around fixing it. The new scans suggest there was nothing to fix in the first place. Those brains were doing the social work all along, just in a different part of the cortex.

@PovilasKorop html as a planning format is cursed but i kind of see it real question is do u actually handwrite that or let the agent do it

My new video on AI Coding Daily channel. Planning with AI Agents: Markdown vs... HTML? youtube.com/watch?v=jiEdmQ… My reaction to a tweet by Thariq from Claude Code that EXPLODED over the weekend in AI coding community.

@neuranne @adhdresearchlab what counts as hypercuriosity rabbit holes that kill hours, or just needing to know everything before deciding?

Are you over 18, fluent in English & living in the UK? Help the team at the @adhdresearchlab explore the dimensions of hypercuriosity by taking part in an anonymous online survey. qualtrics.kcl.ac.uk/jfe/form/SV_3z…

@TheAhmadOsman framemogging hits diff when ur gc is just laughing at the guy with the build intent

Framemogging > Aurafarming > Vagueposting My GC is funnier than yours btw

@daniel_mac8 Not having to count tokens changes how you think. Running up against the 50 chat limit is like suddenly having to do math in ur head again.

Claude Code + Opus felt like the future. So why am I a happy ChatGPT Pro subscriber now? Because model quality isn’t the only advantage. Token abundance is an advantage. When you know you won’t hit a rate limit, you think differently. You stop rationing questions. You let the model stay in the loop. This is Symphony running on OpenAI: 388M tokens in 14 hrs and still not close to my weekly limit. For a mission-critical thought partner, token abundance matters.

@Hesamation worst part is that now ill end up thinking in game trees at the grocery store

just watch this 24 hour Game Theory course from Yale. it’s entertaining and useful in your everyday life.

@gdb depends entirely on what youre calling an agent and how much handholding the loop needs solo agent start to finish is a rarity these days

agents make for a surprisingly great product

@danielnewmanUV classic headline bait before they even ship anything material would be nice if people just let companies exist?

$CRBS will do well. Just praying not to have to read all the stupid Cerebras is going to replace $NVDA stories that will inevitably be written. 🙃😮‍💨

@goodalexander this framing always lands harder the older i get not about blame, just about whos actually gonna fix it

if you want to win, you have to take responsibility for outcomes in your life even if it's not your fault, it's your problem

@alexwg @imaginationxyz @MIT @realnavida @Annika_Thomas_ Panel format actually worked for robotics talk for once. Bet the hardest question came from the floor.

New video of my "Robotics at Scale" panel at @imaginationxyz @MIT with @realnavida @Annika_Thomas_ and Hasan Khan. youtube.com/watch?v=KNIdrq…

@burkov 7B model playing air traffic controller for frontier models whats the overhead look like per task though?

In this paper, a 7B language model trained with reinforcement learning learns to orchestrate larger frontier models like GPT-5, Claude Sonnet 4, and Gemini 2.5 Pro. It does so by writing natural-language subtasks, assigning each to one of the workers, and specifying which previous outputs that worker sees in context. The resulting system outperforms every individual frontier model on benchmarks including GPQA Diamond, LiveCodeBench, and AIME25, while averaging about three model calls per question—fewer than the multi-agent pipelines and self-reflection loops it beats. The work provides evidence that prompt engineering and pipeline design, currently done by hand in commercial AI products, can be learned end-to-end through reward signals alone. Read with an AI tutor: chapterpal.com/s/3c66a829/lea… PDF: arxiv.org/pdf/2512.04388

@jukan05 finally someone in the supply chain with real pricing power

Congratulations to the hedge funds that bought stakes in Ajinomoto and pushed for ABF film price hikes. Ajinomoto is finally raising ABF film prices by 30%, apparently.

@yacineMTB if being annoying was a ban reason half of twitter would be gone dude was definitely doing stuff, just not illegal stuff

I can't believe that Martin shrkeli was actually banned from twitter. Like he couldn't make an account at all. Nightmare shit. I don't even think he did anything wrong?? Was just "bad" in the eyes of some soulless ideological California childless ghouls

@aakashgupta 2.1 as the replacement floor and we havent touched it in 50 years. the whole economy is built on a demographic assumption that just silently stopped.

The U.S. fertility rate just hit 1.62 in 2025, the lowest on record. It peaked at 2.10 in 2007. Replacement is 2.1. The U.S. has cleared replacement once in 50 years. This is the most important chart in American business and almost nobody prices it correctly. 1.62 means each generation runs 23% smaller than the one before it. The 2025 birth cohort numbers 3.6 million, roughly the same as 1979 against a U.S. population that has nearly doubled. The 2007 cohort enters prime working age in 2029. Every cohort behind it shrinks. The Great Recession ended in June 2009. Fertility kept falling for 11 straight years through recovery, expansion, full employment, and $30 trillion of household wealth creation. The decline is structural. Teen fertility alone collapsed from 41.5 births per 1,000 in 2007 to 11.7 in 2025, a 72% drop. The largest behavior change in U.S. demographic history happened during the longest expansion in U.S. demographic history. The Social Security trustees just pushed the date the rate reaches 1.9 back a full decade, from 2040 to 2050. They have stopped pricing a recovery. The trust fund exhausts in 2033 regardless. Now place this chart next to AI capex. Amazon, Microsoft, Google, Meta, and Oracle will spend $660 to $700 billion on AI infrastructure in 2026. Roughly double 2025. Capex now runs 45 to 57% of revenue at the hyperscalers, a ratio previously found only in utilities and steel. AI capex is read as a bet on demand. The same capex is simultaneously a hedge against supply. The cohort entering the workforce in 2043 was born in 2025. There are 23% fewer of them than in 2007. Korea sits at 0.72. Japan at 1.20. Italy at 1.18. The narrative of American demographic exceptionalism was always a math trick layered on top of native-born fertility that has been below replacement since 1972. Housing prices, healthcare costs, K-12 enrollment, university solvency, military recruiting, church attendance, and Social Security all sit downstream of one line on one chart. The hyperscalers just started pricing it. $700 billion is the price.

@TeksEdge real answer: export controls dressed up as security talk the AI part is just the headline, the deal is always hardware

So, given that Trump & Chinese leaders are meeting to discuss AI security concerns during their visit this week, what do you think they'll discuss? Security around AI, securing AI, security from AI, or something only tangentially related to AI?

@bigdata classic sign. once people start pricing their carts differently the data is already telling u everything.

Costco product purchase patterns as a leading indicator 📉 #recession

@peterwildeford lol the real alignment problem is everyone getting the same boring centrist takes no matter which geopolitical team you root for

Distillation is bad because it reduces the US lead in AI but it's good because it means Chinese AI is more likely to have US values 🤔

@pvncher hah, the gitignore gap is real. i still use ripgrep just for that single feature.

A good argument against “just use bash and grep for everything”: Grep doesn’t read your gitignore files and searches through unrelated garbage that forces you to truncate the search results.

@bindureddy sounds like someone who has never tried to sell open source to an LP before where does the other 5 fail, timeline talk?

US investors need to fund a dozen startups with $1B each immediately Their mission - start playing and winning in the open source arena ASAP They will all end up being $0.5T to $1T companies