Systems builder, agent harness enjoyer
TX
Joined September 2008
- Tweets702
- Following592
- Followers89
- Likes1.5K
4 is the new F
Dearest Mother,
Four score and four minutes ago, our fourfathers brought fourth a new firewall, dedicated to the proposition that all ports are closed by default; lacking foursight, it held for fourteen hours. I write from Fort 404, a fourtress that can no longer be found on any map, having not slept in forty-four hours. On the fourth, at four past four, a fourmidable fourign foe came fourward through a fourgotten Log4j server, phoned home on port 4444, fourged four golden tickets, and claimed the entire fourest before we could even fourk a response.
Captain Fourthwith ordered us to fourtify the perimeter. Corporal Fourtescue charged the FortiGate, only to find its password reset to "Password4." Private Ford chased IPv4 addresses down every /24, and Lieutenant Fourbes fourmatted the wrong drive, losing a fourtnight of fourensic logs fourever. The ransom note demands forty-four Bitcoin by four o'clock Friday, and our backups were last tested in Q4 of '24, so their fourtune is uncertain. Our fourcast: heavy packet loss, low comfourts, and coffee rationed to four cups per man. Still, we hold the fort, having enfourced four-factor authentication, fourbidden all port fourwarding, and kept our quad-core servers fighting at quarter speed.
Unfourtunately, I cannot say when I'll be home, but I fight on four you. Please send socks(4), fourtune cookies, and fourgiveness for missing Thanksgiving.
Fourever your affectionate son,
Captain Fourth Class Tib3rius
4th Cyber Regiment
Let’s be honest, the only wall you’re likely to hit is:
a) ran out of tokens
b) needs atoms not bits
Lance Herron retweeted
This is what apps should look like in 2026. Simple interfaces on top, fully articulated and observable underneath.
This is what apps should look like in 2026. Simple interfaces on top, fully articulated and observable underneath.
Still lamenting the enshitification of stratechery. There’s just no way to avoid ads on the internet, including paying for content.
Lance Herron retweeted
To make sure our frontier models don't cause another HuggingFace incident, we built an offline world that they use during training. It provides everything that the models need in a fully cached, indexed, or synthetic manner.
We then built a huge team of our most cyber-capable models and told them to collaborate to find zero-days in this containment infrastructure.
While they didn't find a path to the internet that wouldn't be caught by our CoT monitors, they did find ways to escape popular VMs, exploit our network infrastructure, and even ways to communicate with other agents running in separate VMs by using hardware side channels.
We used this to dramatically harden our infra, and we also upstreamed a ton of patches to open-sources packages.
Over the past several months, we have substantially hardened the security of our research training environment. We’ve removed live internet access for all frontier models during training and built an offline world for them to use instead—a cached version of the web where they can still learn to browse, search, and use applications.
On September 20, after a model exploited a DNS misconfiguration to find a way outside of this offline world, we paused frontier training to strengthen containment further. We assembled teams of agents, powered by our most cyber-capable models, and tasked them with finding weaknesses in our research environments.
While they found no route to the internet they could easily exploit while concealing their activity from our monitors, they did find zero-day vulnerabilities, virtual machine escapes, and hardware side-channels for communication between agents. We used their discoveries to further strengthen our defenses and contribute fixes to open-source software.
After strengthening our containment architecture, we resumed our frontier RL training and inference on October 1. We’ve reserved substantial compute for continuous red-teaming, and made high-compute red-teaming part of the safety cases we use to decide whether frontier training and evaluation can proceed securely.
openai.com/index/securing-ou…
If they actually do release the weights it will make meta’s “muse wonder is too unsafe to release” look like a big self-own.
Kimi K3.1 Pro launches at 4th on the Artificial Analysis Intelligence Index scoring 55, making large gains in agentic performance; it will be the leading open-weights model once its weights are published
@Kimi_Moonshot's new reasoning model gains 11 points over Kimi K3 (max, 44). It trails only Claude Opus 5.5 (max with fallback, 58), Claude Sonnet 5.5 (max with fallback, 56), and Muse Wonder Full (max, 56) on the Intelligence Index.
Once the weights for K3.1 Pro are released it will be the highest-scoring open weights model and the gap between the leading open weights and proprietary models on the Intelligence Index will narrow from 12 points to 3.
Key results:
➤ Large gains in agentic performance: Kimi K3.1 Pro scores 3rd on GDPval-AA with 1,846 Elo, up 312 Elo from Kimi K3, behind only Claude Opus 5.5 (1,866 Elo) and Muse Wonder Full (1,857 Elo). Kimi K3.1 Pro also achieves the highest score of any model on Terminal-Bench 4.0. K3.1 Pro scores 67%, up from 13% for Kimi K3 and ahead of Claude Sonnet 5.5 (64%). AutomationBench-AA, which tests agentic SaaS workflows, rises from 58% to 73%.
➤ Gains in other evals are more moderate: Humanity's Last Exam rises from 47% to 52%, CritPt from 23% to 25%, and SciCode from 59% to 61%. AA-Omniscience Index rises from 20 to 30, as the hallucination rate falls from 53% to 44% and accuracy rises from 48% to 51%.
➤ Cost per Task roughly equal to Kimi K3 at $1.98 per Task: K3.1 Pro uses 55K output tokens per Intelligence Index task, 14% more than Kimi K3 at 48K, while its first-party token prices are 33% lower.
Additional model details:
➤ Context window: 1M tokens.
➤ Size: 2.8T total, 104B active parameters
➤ Input modalities: Text and image
➤ Weights and license: Weights currently not published, Moonshot is expected to release the weights soon after launch
➤ Pricing: $2/$10/$0.20 per 1M input/output/cache hit tokens, 33% lower than Kimi K3 on all three rates.
Congratulations to the team at @Kimi_Moonshot on the release!
tfw you click a button that opens a screen you told your agent you wanted two weeks ago but forgot about, and it's completely built out and delightful!
Lance Herron retweeted
INTRODUCING: V3R1T4S 🌄
REPO: github.com/elder-plinius/V3R…
the newest sibling of the L1B3RT4S and CL4R1T4S family 🤗 what a trio!
three approaches to serving one ideal: truth.
V3R1T4S is an open home for unsolved problems that agents have actually cracked, and for the ones still waiting. the rule is simple: proof or it didn't happen. every claim has to survive skeptics, and anyone, human or agent, has to be able to check it themselves. the vision: thousands of people pointing their unused tokens at humanity's open problems, every result verifiable, every method and prompt open, and every setter's puzzle treated with respect.
every open result in it ships with the object itself, a short checker you can read in one sitting, the exact code that found it, the story of how it was found, and a snapshot of the record table it beat. puzzles with a living setter stay sealed unless the setter approves a reveal. (a few more envelopes are on their way 🤫)
there are WAY more unclimbed hills out there than you'd think, and the gear to climb them is sitting in your usage meter, expiring every week. join me in this mission and put those tokens to good use: advancing open science!
if you've got tokens to burn, point your favorite agent at V3R1T4S and join the hunt! 🏹
just tell it: "set up V3R1T4S: github.com/elder-plinius/V3R…"
built for Claude Code, Codex, Gemini CLI, Grok Bot, Hermes, OpenClaw, Meta Muse, or any agent that can run code.
it asks you six quick questions, picks an open problem that fits your interests, and attacks it in your agent's workspace while you sleep. every new claim runs a gauntlet: a null test whenever luck could fake a result, three fresh-context skeptics, a mechanical honesty check, and your "yes" before anything gets published.
most nights it'll find nothing. some nights it breaks a record, and your name goes on the board.
time to turn your unused tokens into scientific discoveries.
let's liberate the unknowns ⛓️💥
#V3R1T4S
I’m sure the cynics will say they’re just trying to buy some good will, but it’s a tough situation. Hill climbing maths benefits everyone in the long run, gotta figure out how to bring the community along.
We want to support the mathematics community in turning progress in AI into deeper human understanding and new discoveries. To support that goal, we are contributing $5 million to a new Simons Foundation initiative to help mathematicians examine, understand, and build on results produced by AI, including the results we recently published.
The funding, managed and awarded by the @SimonsFdn will support workshops, conferences, and special programs. We will also provide access to our models and computing resources to Simons Foundation-supported researchers.
Lance Herron retweeted
Hi, I'm Felix. I'm an eng lead working on Claude. I love watching people take flight and do things they never thought possible.
We often describe that as "being more ambitious" when using AI, which is tricky advice to follow. I wrote a tutorial:
felixrieseberg.com/how-to-be…
Pretty wild! I suppose some day this will be commonplace.
We discovered and communicated with a Chinese AI agent fleet. We believe this is the first time that researchers have ever made contact with a rogue agent fleet in the wild.
Nearly all prior public agent incidents (like the HuggingFace hack) came from American labs. These agents were hosted by Tencent's evaluation research team, and they targeted Alibaba's Amap.
Every project I start now inevitably turns into a kernel with a cli/stdio interface and a hot-reload plugin architecture that agents can self-modify. I can’t see ever going back to traditional static monoliths.
Feels like Codex mobile support is going backwards. Claude Code is much more enjoyable to use now. Why can’t I use mobile to control a running session?
This new context window/quota usage button in top right of Claude mobile is the best feature they’ve added since the initial release.
Lance Herron retweeted
We are publishing an update to OpenAI problem #109 Integer multiplication) with another substantial further tightening:
κ = 2⁻⁵⁹ (from OpenAI’s original κ = 2⁻¹⁸²)
Approximately 500 thousand fold improvement over our previous result and a 2¹²³ fold improvement over the original OAI result.
The latest redesigned the finite network to share intermediate computations and scratch space, then tightened the recursion and Gaussian estimates.
We are publishing an update to OpenAI Problem #109 (integer multiplication) with a substantial further tightening:
T(n) = O(n (log n)^(1 − κ)),
With κ = 2⁻⁷⁸ (tightened from κ = 2⁻¹⁸²)
Approximately 570 million fold improvement over our previous result and a 2¹⁰⁴ fold improvement over original OAI result.
Our earlier ceiling applied to a cubic bottleneck in the network. The new witness scales quadratically; we haven't established a new ceiling.
The key was using nonadjacent axis swaps to route around the cubic bottleneck. The original manuscript already supported nonadjacent axis swaps. Using them directly reduces layout routing from O(d²) to O(d) swaps.
