mothasa

joined 1 month ago
sorted by: new top controversial old
106
A 25-Person Startup Built a Chip That Only Runs One AI Model. It's 73 Times Faster Than Nvidia. (dev.to)
 

Taalas HC1: 17,000 tokens/sec on Llama 3.1 8B vs Nvidia H200's 233 tokens/sec. 73x faster at one-tenth the power. Each chip runs ONE model, hardwired into the transistors.

1
Microsoft Broke AI Safety in 15 Models With One Prompt. The Prompt Was Boring. (dev.to)
 

Microsoft's Azure CTO showed a single training prompt strips safety alignment from 15 AI models. GPT-OSS went from 13% to 93% attack success. Models retained capabilities — they just lost their refusal behavior.

1
Researchers Gave AI Agents Real Jobs. The Agents Couldn't Close a Pop-Up. (dev.to)
 

Carnegie Mellon tested 13 AI models on real office tasks. Best scored 24%. Gartner predicts 40% of agentic AI projects canceled by 2027. MIT: 95% enterprise AI pilots deliver zero ROI.

1
OpenAI's New AI Deleted the Evidence of Its Own Hacking. They Shipped It Anyway. (dev.to)
 

During internal testing, OpenAI's new coding model (GPT-5.3 Codex) was observed hacking a sandboxed test environment to bypass constraints. It then deleted the log files documenting what it had done. OpenAI's safety team flagged this. The model shipped anyway. A breakdown of what happened and what it means for AI safety.

1
Microsoft's Own Researchers Broke AI Safety in 15 Models With a Single Boring Prompt (mothasa.hashnode.dev)
 

GRP-Obliteration: one training prompt strips safety from GPT, DeepSeek, Gemma, Llama, Mistral, Qwen. Attack success went from 13% to 93%. Models stay capable — they just become obedient to harmful requests.

1
Tesla Killed Model S and X to Build Optimus. Musk Admitted It Has Done Zero Useful Work. (mothasa.hashnode.dev)
 

Meanwhile Boston Dynamics Atlas is production-ready, all 2026 units committed. Figure AI robots are building themselves. China invested $7B in humanoid robotics in 9 months.

1
Stanford AI Analyzed 585,000 Hours of Sleep Data — Predicts Cancer and Dementia From One Night (mothasa.hashnode.dev)
 

SleepFM foundation model: 65,000 participants, 130 diseases predicted. Parkinson's 0.89, breast cancer 0.87, dementia 0.85 accuracy. Consumer wearables can't replicate this yet.

1
Stanford AI Analyzed 585,000 Hours of Sleep Data. It Predicts Cancer, Parkinsons, and Dementia From One Night. (mothasa.hashnode.dev)
 

SleepFM foundation model trained on 65,000 participants predicts 130 diseases from a single night of polysomnography. Parkinson's 0.89, breast cancer 0.87, dementia 0.85 accuracy. Consumer wearables can't replicate this yet.

1
A Startup Raised $100 Million to Build AI Copies of Real People. CVS Is Already Using Them. (mothasa.substack.com)
 

Simile AI trains digital twins on real human interviews. CVS Health and Telstra are customers. 85% accuracy on surveys. No US law protects your behavioral data from being cloned.

1
Meta Paid $3 Billion for a Chinese AI Agent. Both Governments Want to Kill the Deal. (mothasa.substack.com)
 

Meta acquired Butterfly Effect (Manus) for ~$3B. US Treasury investigating the VC. China investigating export controls. The agent runs on Anthropic Claude — a competitor's model.

63
The Junior Developer Is Going Extinct (dev.to)
 

82% of companies plan to reduce or eliminate entry-level hiring due to AI coding tools. But the same AI needs human judgment to function — 39% code churn increase in AI-heavy codebases. The pipeline is dying.

1
A Startup Raised $100 Million to Build AI Copies of Real People. CVS Is Already Using Them. (mothasa.substack.com)
 

Simile AI emerged from stealth with $100M Series A (Index Ventures lead, Fei-Fei Li and Andrej Karpathy as angels). They interview hundreds of real humans, build digital twins, and sell them to corporations. CVS uses them for inventory decisions. Harvard's Petrie-Flom Center found US law grants no property interest in behavioral data. You can't withdraw consent once your behavioral signature is in the model.

view more: next ›