`torch.load()` doesn't read a file. It executes whatever Python the file tells it to. The scanner HF uses to catch this (PickleScan) has 3 published bypasses, all CVSS 9.3. brysgo.com/post/2026-08-1…
Researchers fine-tuned GPT-4o on a simple prose task. It reproduced 460 verbatim words of copyrighted text. Alignment didn't erase that data — it suppressed it. Fine-tuning turned the key. Code & Cosmos episode:
A 60-year-old Erdős conjecture fell to an AI in 80 minutes. Then Terence Tao spent days just figuring out what the proof actually meant. The bottleneck in math didn't disappear — it moved from proving to auditing.
Most enterprise AI budget goes to sales and marketing. MIT's data says the real returns are in accounts payable and audit — invoice processing cost dropping from ~$13 to ~$3 a pop. The boring, rules-based stuff is what actually pays off. Wrote up why: brysgo.com/post/2026-08-0…
Half of enterprise AI budgets chase sales and marketing. The real returns hide in invoice matching and audits — up to 1,800% ROI in 3 years. New Code & Cosmos on why the boring agents are winning:
