The full catalogue of scrollytelling pieces, visual essays, and data reports from Vizmaya Labs — geopolitics, technology, demographics, and the slow variables that explain the fast headlines.
Tesla has driven about 2 million unsupervised miles ever. Waymo drives about 4 million every week. The market still pays 160x profit for Tesla.
A small AI swarm tried 734 programs in an hour vs 158 for its best rival, and its 8B model scored 0.804 vs 0.774 for rivals on a 32B model.
A 2026 paper adds four steps and human sign-offs to AI software design. Its own test found different scores on the same input.
In a 2024 PNAS Nexus study, readers of plain GPT-4 science summaries explained the research better (0.92 vs 0.72) but rated the scientist as less smart.
In a 2024 PNAS Nexus study, readers explained research better after plain GPT-4 summaries (0.92 vs 0.72), but trust gains were mixed.
Humans lead robots 9.66 billion to 1.0 billion in 2050, even on a bullish forecast. A new paper shows what it would take for robots to catch up.
AIRoA-0 learned two-handed robot tasks from 16,151 hours of handheld human demos and scored 78% vs 27% for the best rival on ten tasks at an unseen site.
On a 16B model, MMLU-Pro rose from 23.86 to 30.86 after its top layers reused earlier layers' notes, with a third less cache memory.
The GPU busy number only says a task was running, not how much of the chip it used. A new paper times the hidden steps. What holds up, and what does not.
A 2026 paper trains a small AI guesser on its own guesses. Training ran 16.09x faster and guesses lasted longer, but gains over the best rival were small.
A standard AI model wrote a valid trip plan for 10% of problems. Put the rules inside the decoder and every plan was valid, up to 1.43x faster.
An AI agent loop took one H100 from 184 to 416 tokens per second without changing answers, but one trick made batches 13% slower.
Authors report a 27B AI scores 80.2 on a computer-use test. Its real edge: a terminal took one hard task from 6.6 to 43.9 in fewer steps.
A preprint reports an AI that learns where to point its magnifying glass scores 71.6% on seven vision tests. Unreplicated, no code yet.
A Tsinghua preprint reports that four attention heads decide where a small vision AI looks. Training them lifts accuracy 74.1% to 76.6%.
A new paper says home AI stalls because reading and writing want opposite things. Its own test credits plain plumbing: up to 2.24x in decode.
A new paper took down a fixed memory wall inside a GPU and ran 90 chats at once instead of 34. What it shows, what it costs, and what is unproven.
A new study found AI agents waste compute after a tool call. A small controller cut 24.6% to 43.1% of the work with accuracy about level.
At the same 32,768 training tasks, Apple's SynthAPI recipe with more apps, more worlds and more clutter beat a 'more tasks' recipe in 12 of 12 comparisons.
Across 42 editions of AI-buildout news, countries with 25.8% of world GDP and 46.2% of people got 2% of country tags. The US alone got 48.9%.
A lone AI coder stalled at 0.51× expert speed on a GPU task. A team that rebuilt itself reached 0.74×, then 1.21×, says a Berkeley preprint.
By step 3 of a hard question, 70% of an AI search tool's notes hold its own wrong passage. Training on its own mistakes lifts full answers from 44.6 to 50.5.
Alibaba's KARMA lets an AI ad agent edit its own rulebook, one tested rule at a time. Average score on unseen cases rose from 30.19 to 39.15 of 100.
Phone flash answers in 40 µs, memory in 30 ns. A new design, tested only in simulation, runs an AI agent team 21.4x faster. How it works and where it breaks.
A speed trick makes AI answer up to 4.3 times faster for one user, but the energy saving fades as users join and can turn into a loss.
Google's Finnish subsidiary felled about 525 hectares (1,300 acres) before its environmental assessments. The wolf 'invasion' rests on one villager's quote.
Bell, Curie and Einstein each beat a written record an AI would have trusted. Why judgment, not prompting, is the skill that matters.
Anthropic says juniors fell from 27% to 9% of its tech hires. Fed and BLS data show no national break yet: ages 20 to 24 unemployment is 8.0% vs 9.2% a year ago.
Amex will back your shopping agent’s mistakes, right after you’ve tried returning them to the merchant. Meanwhile, agent checkout is still measured in basis points.
SK hynix gave up 8 points of AI memory share in one quarter while its total sales rose 51%. Here is how the market works, from the ground up.
How a strait reprices a GPU hour.
OpenAI is building 9 gigawatts of AI infrastructure across seven US sites. Three years in, only one is running. The map of where Stargate is being built is really a map of where you can plug in fastest.
Memory, carbide, EDA and AI clusters. China is announcing all of them at once.
Materials, power, packaging, cooling and land. Nine bets on where AI infrastructure gets built.