Ryan Carson runs 15 Devin agents for his family-law SaaS. His honest take on what that actually looks like might surprise you.
Every agentic AI security tool works both ways, the same capability defends and attacks. Ben Thompson argues incentives therefore tilt toward offense, an imbalance that cements incumbents into defensive postures while underfunded attackers iterate fast. Defenders who can't use the best models are left patching with the enemy's open weights. That asymmetry, he says, is what favors attackers and fuels startups over time.
ChatGPT Ad Library claims to be the first public registry of sponsored ads inside ChatGPT, tracking 11,103 advertisers, 415,289 placements and 43,410 creatives across 970 niches. Each ad is traced to the exact prompt that triggered it. OpenAI publishes no ad transparency center for its growing ad business, so this crowdsourced catalog is filling that gap for marketers who want to spy on competitor creative and lift CTR.
This repo builds on Karpathy's microgpt to show LLM watermarking at its most minimal: a secret-keyed Gumbel sampler swaps in for random sampling without touching the model's token probabilities. Generated text stays indistinguishable from normal output, yet the detector flags it at p ≈ 6.8×10⁻⁵³ while a wrong key lands at chance. It's proof the watermarks modelers actually ship are just a few nudged decisions, no frameworks required.
Five-time founder Ryan Carson is running Untangle, a family-law SaaS, on $20,000/month of concurrent Devin agents, engineering, customer success via a "Watchdog" monitor, even his PR pipeline. His honest take: more AI output is the wrong goal. Managing 15 agents comes down to folders, P0 threads, and a paper list, not a dashboard.
You’re caught up.
That was all 4 stories in the 25 August issue.