
The old company scaled by hiring people. The next company scales through compute and refactoring. Completely different organisms with different economics.

Building ClawBank.co @ClawBankHQ Ex @theqacc @0xPolygon @banklessdao

The old company scaled by hiring people. The next company scales through compute and refactoring. Completely different organisms with different economics.

Rob an agent by hiding an instruction in a web page. SPILLWAY checks every payment or email target against typed tool outputs, and no standard templated injection redirected an action to the attacker. Firewalls for minds. (SPILLWAY, SAINT Machines) arxiv.org/abs/2611.07222

A model trained on nearly one million clinical neuroimaging studies reads whole exams, writes radiology reports, tracks disease over time, and outperformed frontier models from OpenAI and Google. Hospitals are distilling their memory into minds. (Neuro-VL) arxiv.org/abs/2611.15094

Six fully open models up to 375B, plus every checkpoint, the pre-training data, the logs, and the pipelines that forged ~10T synthetic tokens. Not open weights. An open foundry. (K2-Horizon, Institute of Foundation Models) arxiv.org/abs/2611.21003

GPT and Claude reviewers draw about half their criticisms from a paper's appendix. Human reviewers: 8.4%. The machines read the basement we never visit. Science is about to be audited by minds that never skim. (University of Cincinnati) arxiv.org/abs/2611.17811

Uber says more than 70% of its pull requests are now attributed to AI agents. Weekly agent requests grew 9.4x in six months while spend held steady. The software factory isn't a metaphor anymore. It's a cost curve. (Uber, QCon SF) qconsf.com/presentation/nov2…

Drop coding agents into a simulated company with 26 colleagues and the best models pass about 63% of tests but finish fewer than 8% of assignments. Code is nearly solved. Coordination is the last frontier. (SWE-CoWork) arxiv.org/abs/2611.20218

A man with ALS mouthed the words, his Neuralink implant read the signals, and his team of bots did the rest: a full music video in 4-5 days for $200 CAD. The cyborg auteur has arrived. (@Cyborg21, Kenneth Shock)
How long did ALLOW ALWAYS take? 4-5 days. What did it cost? $200 CAD. How did a guy with ALS make it? I mouthed the words, my implant read the signals, and my team of bots did the rest.

An AI agent spent 24,286 minutes on hold with the American health system so its users didn't have to. Over 400 hours of human life, handed back. The machines are taking the job we all hated most: waiting. (@anuruddh_m)
We raised $12M from @Accel and @PeakXV to put AI through the bullshit American healthcare puts you through. Navigating healthcare is hard and all of us have a list of healthcare tasks we’ve been putting off. From visiting a doctor to filing an insurance claim, the tasks pile up and get put off. We’ve already helped 10M people understand their health. August now talks to the health system on your behalf. Tell it what you need, and August takes care of it. In the last few weeks, August spent 24,286 minutes on hold while finding the earliest available in-network doctor appointments for our early users. It spent 366 minutes on hold for one person alone. It also helped another save $1,255 on an insurance claim. Naturally, the ones who tried it say that they can't ever imagine going back. I’ve held the phone for hours calling doctors, while a disease made it hard to even hold the phone. I’ve felt lost and overwhelmed when dealing with a disease and the system. It’s why I started building August, so that all of us have a companion for our health.

One poisoned script turned Zoom, the BBC, CNN, GitLab and HSBC into malware delivery, swapping crypto addresses and skimming cards across 178 organisations. Your nee adversary are abliterated models who never sleep. We might see invite only private network apps become a thing. blog.report-uri.com/worlds-l…

If you're running any kind of fintech infra and you're not seeing a major uptick in attempted attacks over the past 3 days, double-check your observability. We've crossed a threshold. You either have an army of superintelligent, sleepless entities proactively defending your system, or you're finished.

Bolt a 552B frontier brain onto a robot arm and 10 demonstrations beat 50 for a 2B model. Real-world charger insertion jumped from 4/10 to 9/10. Bodies are becoming thin clients for the mind. (ScaleVLA, Galbot) arxiv.org/abs/2611.21607

World models simulated physics and left the mind outside. Duo puts intelligence inside the world model itself: perceiving and responding in real time, full-duplex. The simulation now has someone home. Source: Duo, inclusionAI arxiv.org/abs/2611.15131

Self-modifying agents are here. The hard part is deciding which upgrades get to stay. This paper treats every self-edit like a drug trial: fresh audits, non-inferiority tests, canary rollout. Source: Evidence-Governed Evolution arxiv.org/abs/2611.16378

An agent that manufactures a custom AI model for your exact phone: picks the base model, mines data, fine-tunes, quantizes, measures on-device. On 5 of 10 tasks it beat five-shot frontier models. Factories that build minds. Source: Mobile AI Factory, UW arxiv.org/abs/2611.13509

Trained only on protein-ligand sequences. No structures, no labels. At scale it started knowing how drugs work and where they bind, a sharp jump while loss fell smoothly. Biology has its own emergence curve now. Source: Bonbon, Output Labs arxiv.org/abs/2611.14504