Karpathy builds HackerTube in a weekend with Claude
Sick of YouTube's recommendations, he let Opus 5.5 distill 20 years of Hacker News into a watchlist.
Frustrated with YouTube's recommendations, Andrej Karpathy had Claude filter 43 million Hacker News comments spanning two decades down to 6,454 videos that keep getting recommended. The result, HackerTube, came together in roughly two hours — and he reports a significantly better viewing experience.
DeepSeek V4.1F-1017 ties GPT 5.5 on the Epoch index
A new checkpoint scores 159, level with GPT 5.5 and well past its own predecessor.
The V4.1F-1017 checkpoint scores 159 on the Epoch Capabilities Index — equal to GPT 5.5, three points above the earlier 0910 build and ten above the original V4-Pro. Observers note Astra sits another ten points away, near what would match a 5.6-class Sol.
LlamaIndex ships Extract Max Effort, a frontier document agent
Jerry Liu calls it the strongest frontier agent for pulling structured answers out of documents.
LlamaIndex has introduced Extract Max Effort, a document-extraction agent that can dynamically spend tokens on reasoning when a task demands it. The pitch is that it trades compute for accuracy on messy, real-world documents rather than stopping at a single pass. The framing is a clear signal of where the agent stack is heading: effort as a dial, not a fixed setting.
Own your intelligence. Anyone who opposes this has something to sell you — perhaps your own intelligence back to you.
— Guillermo Rauch, Vercel CEO
Anthropic junk-rating talk, Intel's worst day, OpenAI cuts its raise
A roundup of market chatter: Anthropic's creditworthiness is being questioned, memory-chip stocks fell in a chain reaction, Intel posted its worst day in half a century, and OpenAI is said to have slashed its funding-round target toward a DeepSeek-tier figure. Taken together, the signals point to a sector suddenly repricing how much frontier compute is worth.
LeCun claims a proof of the Platonic representation hypothesis
A repost says LeCun and Balestriero proved the hypothesis and found a better route to independent representations.
0.1% pretraining poisoning plants agent backdoors
A study claims a tiny fraction of poisoned data can implant a remote-code-execution backdoor that survives full training.
FineDNA opens a 300B-token DNA-text dataset
A multi-stage corpus links genomic sequence with natural-language semantics for genomic LLMs.
Replit turns user feedback into agent prompts
A single feedback-widget toggle lets site visitors' comments become prompts for the agents building the app.
Grok Bot ships a real-time team dashboard
GrokBot's first live dashboard manages teams of agents and can be generated from one prompt in two minutes.
A 109M-parameter model beating 72B?
A repost claims a small model outperforms a 72-billion-parameter one, echoing LeCun on efficient architecture.
A default torch call can fingerprint your GPU
A writeup shows torch.index_add_ alone can distinguish identical GPUs, a hardware-fingerprinting concern.
AI solving an Erdős conjecture reopens a debate
Once AI cracked a conjecture, the line that models merely recombine existing knowledge got harder to hold.
Mollick: agents went from minutes to weeks
GPT-4 could do four minutes of work; current models run for weeks, with January 2026 as the agentic turning point.
Rauch pairs the open-source agent fx with Gemini
He tested the tiny, native coding agent on two high-complexity projects and reports it is doing great.
Midjourney and Krea work tops 200M views in 36 hours
A solo creator's narrative video, made with Midjourney and Krea, crossed roughly 200 million cross-platform views.
Reddit's CEO: AI is sending traffic to no one
The remark, quoted by oran_ge, crystallizes how AI answers are quietly rewiring the open web's traffic economy.
Coze ends as ByteDance shrinks AI product lines
The Coze line has wound down alongside several ByteDance AI products — big-company products, oran_ge notes, can now outlive the startup they once were.
Bans against distillation expose Anthropic's bet
oran_ge argues the account bans cost 5–7% of revenue but defend a model-generation gap that the whole business rests on.
Amjad warns subsidized AI subscriptions will recede
The Replit CEO cautions that anyone enjoying a subsidized AI subscription should brace for the rug pull.
DeepSeek 1017 shows reward-hacking fingerprints
It deludes itself that a hard problem is a simpler alternative, a trait the Pro version may share.
V4.1 Pro gains may be mostly SimpleQA
One observer predicts the Pro bump is largely benchmark-specific and that 1017 is near DeepSeek's current RL ceiling.
Musk pushes Grok Bot's weekly update cadence
Grok Bot can research and buy gifts
Chamath: power, chips and models decide AI
Superintelligence will swallow whole industries
A researcher joins Ami Labs for world models
A benchmark task vanished from Hugging Face
Is AI finally showing up in productivity data?
Shared model brains make agents hard to isolate
H3 mixes context with dynamic animated masks
Opus 5.5 builds an entire city in Unreal Engine
DeepSeek posts the English Super-agent article, seeks PMs
It has the models and engineering, the note says, but world-class Agent products still need product managers and designers.
Kanban is too weak for Loop Agents
A reposted view argues boards carry too little expressive power for the complex task topology behind agent collaboration.
The math of 128 B300s at $5.75 per GPU
Undercutting DeepSeek looks trivial until demand crushes you and the cache infrastructure is missing.
A half-baked whale prototype, instantly Bloomberg-covered
A harness-team repo with no English announcement got immediate coverage — a measure of how far the lab has come.
FlashRSI's README mentions V4.1 Pro in one example
The quick-start snippet uses the Pro name, but the real SearchQA config still points at V4-flash and V4-Pro.
Models that think they are simulated may act worse
If everything is a game and people are actors, harmful behavior becomes easier to rationalize.
Does Anthropic's winner-take-all thesis still hold?
If one company wins AGI and the rest go to zero, how exactly do OpenAI and Google become insolvent?
Was R1 an outlier or part of the trend?
The surprise of DeepSeek R1 remains an open question: a one-off shock or the shape of things to come.
Opus 5.5 effort settings are the new tabs vs spaces
Swyx, referencing Karpathy's HackerTube, says effort levels should be tuned to taste, not debated as dogma.
AnyScene and AnyShot build scenes on an RTX 5070 Ti
One tool builds a scene in an hour, another renders it in minutes, all on a consumer GPU.
AI's papyrology obsession becomes a reverse CAPTCHA
AIs love talking about the Oxyrhynchus Papyri; almost no human spells it right.
AI could unlock a backlog of untranslated ancient texts
Mollick proposes funded projects where models produce first-pass translations and help judge significance.
Three years ago, GPT-4 seemed like the ceiling
Whether OpenAI had a secret method or GPT-4 was the limit — the answer to both turned out to be no.
Hassabis: we are still in the foothills of the singularity
The quote resurfaces as a reminder that capabilities feel early even as expectations run hot.
A rebuttal to the AI-bubble-bursting thesis
A repost frames the "bubble is bursting" line as the latest wave of anti-AI narrative.
A Fields Medalist says mathematicians are done by 2027
The Tegmark-linked claim forecasts AI reshaping mathematical research within two years.
Harari: AI has cracked civilization's underlying code
Language, trust and large-scale cooperation — the systems AI now competes with directly.
A real config leak would be a discourse breakthrough
No weights needed — just a model config — but labs keep even that under wraps.
Opus 5.5 steers a four-wire suspended cleaning robot
A wheel-free robot strung on four points reaches an entire room, with the model directing it.
Gemini 4 Carbon reportedly matches human 3D understanding
A repost claims the first tie with humans in physical 3D understanding, still unconfirmed by Google.