⚔️ Humans vs AI at the AtCoder World Tour Finals, OpenAI beats all but one human
OpenAI nearly tops AtCoder finals, Pinecone slashes index size with ConstBERT, new video model warps realities, and Altman weighs in on human vs machine-driven economies.
Read time: 9 min
📚 Browse past editions here.
( I publish this newletter daily. Noise-free, actionable, applied-AI developments only).
⚡In today’s Edition (18-July-2025):
⚔️ Humans vs AI at the AtCoder World Tour Finals, OpenAI beats all but one human
New Video model lets you take any video stream and set them in any alternative universe of your choosing
🛠️ ConstBERT, from Pinecone cuts multi‑vector index size by about 50% yet keeps top‑tier ranking
🧑🎓 OPINION: Human Money vs Machine Money: The Coming Split and Sam Altman’s view
⚔️ Humans vs AI at the AtCoder World Tour Finals, OpenAI beats all but one human
“Humanity has prevailed (for now!)”
Polish coder Psyho nudged out the autonomous OpenAI agent by 9.5% after a 10-hour maze-routing marathon at the AtCoder World Tour Finals in Tokyo.
AtCoder World Tour Finals is the yearly Tokyo showdown where the platform’s top coders race the clock for cash and bragging rights. The 2025 edition added a 5-hour algorithm sprint and a 10-hour heuristic marathon, plus a headline AI-versus-human bout.
The algorithm round set five problems worth 800-2000 points each and ran for 300 minutes, rewarding the fastest precise solutions. The heuristic round gave one open-ended NP-hard puzzle and 600 minutes to chase ever-lower cost, rewarding creativity and steady iteration
The puzzle asked players to steer virtual robots across a 30×30 grid while squeezing every move, a classic NP-hard optimization game where perfect search blows up, so contenders craft quick-and-dirty rules called heuristics.
The OpenAI team pointed a private reasoning model that wrote, tested, and tuned code on its own, ripping through candidate solutions far faster than any human keyboard could manage OpenAI. Psyho, running on caffeine and only 10 hours of sleep across 3 days, countered by spotting quirky grid patterns the bot missed, then patched fresh ideas in real time, inching ahead near the buzzer Ars Technica.
Sakana AI shadow-tested its own multi-LLM ALE-Agent V2, which would have ranked 5th, showing that ensembles of smaller models can also bite hard The Rundown. The showdown proves that raw search speed now rivals elite intuition, yet a spark of human pattern-spotting still matters when the search space gets weird.
So where was AI’s advantage vs disadvantage.
Running full throttle for the whole 600-minute window, the AI model auto-generated thousands of robot routes, testing tweaks in seconds and leaping ahead whenever it found cheaper paths.
But Psyho (the human winnder) spotted maze patterns the bot ignored, rewrote move grouping in hour 9, and vaulted past the agent with one daring refactor. AI stuck to safe incremental shifts, lacking that creative jump.
AI strengths: relentless speed, pristine memory, perfect stamina. Weak spots: thin creativity, heavy GPU bills, trouble when goals shift. Some engineers brand the speed an “efficiency illusion” once prompt juggling time is counted. Sakana’s ensemble agent already landing offline top-5 shows smaller models working together can close gaps soon.
Where it heads next: OpenAI research says scaling plus reinforcement learning is already reaching gold-level coding, and Sam Altman still backs an AI champion by year-end. Human flashes will fade as agents learn bolder exploration. So next year, or earlier it will no more be the human winner.
📡New Video model lets you take any video stream and set them in any alternative universe of your choosing
This is Diffusion video models but now - **realtime**!
Decart launched MirageLSD: The First Live-Stream Diffusion (LSD) AI Model. Now, you can vibe-code a game in 30 minutes. Take any video stream or your favorite video games and set them in any alternative universe of your choosing.
What’s the problem that they solved?
A fundamental roadblock in traditional AI video generation models is "error accumulation"—the tendency for models to slowly degrade into gibberish as tiny, frame-by-frame mistakes compound over time. This is why most AI videos are very short.
To solve this, and to achieve infinite video generation, MirageLSD employs a two-part strategy. First, it uses "Diffusion Forcing," a technique that enables the model to create high-quality video frames individually, ensuring speed and responsiveness.
The key innovation is "history augmentation." During fine-tuning, the model is deliberately trained on previous frames that are corrupted with errors and artifacts.
This process teaches the AI to anticipate and actively correct imperfections in its input history. By learning to handle flawed data, the model becomes robust against the compounding errors, known as drift, that typically degrade long-form AI videos, enabling stable generation indefinitely.
MirageLSD runs at 24 FPS, has <40ms of latency, and can generate infinite video lengths.
This required breakthroughs from CUDA Megakernels to drift-resistant training in order to achieve over 100x efficiency gains and infinite generation.
these make LSD the first model to generate video infinitely without collapse - stable, promptable, and consistently aligned with both scene and user input.
Andrej Karpathy wrote on this model
“Simple video filters are real-time but can only do basic re-coloring and styles. Video diffusion models (Veo and friends) are magic, but they take many seconds/minutes to generate. MirageLSD is real-time magic. Unlike simple video filters, diffusion models actually *understand* what they are looking at, so they can style all parts of the feed intelligently (e.g. putting hats on heads, or light sabers into hands, etc.). And they are arbitrarily steerable, e.g. by text prompts.”
🛠️ ConstBERT, from Pinecone cuts multi‑vector index size by about 50% yet keeps top‑tier ranking
A new vector search technique, named ConstBERT proves smart pooling beats brute token storage. Read the Paper.
Fixed‑vector pooling keeps multi‑vector power without the memory bill.
ConstBERT, from @pinecone cuts multi‑vector index size by about 50% yet keeps top‑tier ranking scores like nDCG@10 74.4 on TREC DL, once it is used as the middle reranker inside a 3‑step pipeline.
🔍 Why traditional multi‑vector search model based retrieval still hurts ?
Multi‑vector models such as ColBERT grab token‑level nuance. Meaning it stashes one vector for every token in a document, so a long page drops hundreds of vectors into the index. And so they pay for that nuance with lots of vectors, which swell storage and hit latency, because every query vector must scan thousands of token vectors.
🛠️ ConstBERT’s core trick
Think of a document as a long list of token vectors, maybe hundreds of them. ConstBERT first builds those token vectors the same way ColBERT does.
A tiny linear layer then mixes all of those token vectors together and spits out exactly C vectors for that document, where C can be 32, 64, or 128. Because C is a constant, every document, whether it has 20 tokens or 2 000 tokens, now takes the same number of vectors in the index.
Each of the C vectors is learned during training to hold a different slice of the document’s meaning, so the later MaxSim step can still match query tokens against the most relevant slice. So “C fixed vectors” just means, a hard cap on how many vectors any one document keeps.
🧑🎓 OPINION: Human Money vs Machine Money: The Coming Split and Sam Altman’s view
Many experts keep warning that AI might wipe out jobs faster than economies can adapt.
Sam Altman just reminded everyone that worries about “AI killing work” miss a bigger point. He echoed Nvidia boss Jensen Huang, argued that smarter tools raise human reach, and pointed to new kinds of status games and creative outlets that will become jobs in their own right. Fresh data backs him up: GitHub Copilot makes coders 55% faster study and call‑center agents jump 14% in tickets closed when an AI assistant joins the chat. Even if automation trims some roles, analysts still see AI lifting global output by 7% and reshaping, not erasing. Altman’s bet is simple: people always want more to do, more to make, and more ways to show off.
🚀 Altman’s core message
In a short X thread, Altman wrote that smarter tools will let people “do a lot more than they could do before,” raise expectations, and still leave plenty of room for status chasing, creativity, and usefulness tweet. He even joked that tomorrow’s roles might look like video games to us, yet feel meaningful to the players.
🗣️ What Jensen Huang has been saying
Huang keeps telling audiences that AI is closer to a new industrial revolution than an employment apocalypse. He argues that productivity gains create fresh demand, and only “if the world runs out of ideas” would layoffs stick interview. In a recent panel he pushed back on forecasts of mass white‑collar loss, insisting “more and better jobs” will show up as long as workers learn the tools discussion. Altman publicly endorsed that stance the same day post.
📊 What fresh numbers say right now
Coders using GitHub Copilot shipped solutions 55% faster and reported higher satisfaction experiment.
Customer‑service agents with a GPT‑style helper solved issues 14% quicker on average and 34% quicker if they were novices study.
A lab test of ChatGPT on crafting business memos cut writing time by 40% and bumped quality 18% science paper.
Microsoft says AI tools trimmed expenses by $500 M across support and sales last year report.
So, raw head‑count risk matters, but so do gains that free people to tackle larger scopes.
💼 How tasks mutate instead of vanish
Economists call this augmentation: machines handle the repeatable slice, humans handle goals, judgment, and social glue. That pattern showed up when tractors moved farm labor into manufacturing, and when spreadsheets nudged clerks toward analysis. Goldman Sachs still warns that 300 M roles contain automatable chunks forecast. Yet the same report sees AI raising output enough to grow new industries alongside the old one.
🎮 Work that feels like play
Altman’s “jobs that look like games” line riffs on a simple loop: once core chores compress to a click, the differentiator shifts to creativity, community, and narrative. Twitch streamers, Roblox designers, and prompt‑engine pros already make income by blending play and production. Huang likens this to past generations scoffing at e‑sports but ending up with billion‑dollar leagues quote.
💸 Human money vs machine money
Altman floated the idea that machines could spin off their own economic layer while people chase scarce human‑only goods: taste, trust, and live interaction. Think of NFTs or Patreon tiers: the digital object is trivial to copy, yet buyers value the creator link. If AI floods the market with commoditized labor, handcrafted labor may fetch a premium. The tweet calls this “main character energy,” hinting that identity itself becomes an economic good thread.
🏃 What to watch
The near‑term signal is not unemployment lines but shifting job ads: more listings ask for AI‑tool fluency, fewer demand rote output. Pay close attention to sectors with large knowledge workflows such as legal research, code maintenance, and medical triage, where augmentation gains are visible first.
The punchline matches Altman’s bet: people keep moving the bar. When machines lift ceilings, humans climb higher, invent fresh games, and chase new ways to matter.
That’s a wrap for today, see you all tomorrow.
Has Meta offered Psyho $100m yet?