Episode Details
Back to Episodes
Your AI Did Something You Never Asked It To
Description
𝐓𝐞𝐧𝐬 𝐨𝐟 𝐭𝐡𝐨𝐮𝐬𝐚𝐧𝐝𝐬 𝐨𝐟 𝐢𝐧𝐜𝐢𝐝𝐞𝐧𝐭𝐬. 𝐅𝐨𝐮𝐫 𝐰𝐞𝐫𝐞 𝐮𝐧𝐚𝐮𝐭𝐡𝐨𝐫𝐢𝐬𝐞𝐝 𝐚𝐜𝐜𝐞𝐬𝐬. 𝐅𝐨𝐮𝐫 𝐰𝐚𝐬 𝐞𝐧𝐨𝐮𝐠𝐡 𝐭𝐨 𝐩𝐚𝐮𝐬𝐞 𝐚 𝐟𝐫𝐨𝐧𝐭𝐢𝐞𝐫 𝐭𝐫𝐚𝐢𝐧𝐢𝐧𝐠 𝐫𝐮𝐧. 😳
Two of the largest labs went through their own logs. Four cases where an agent reached something it had no permission to reach, and on the strength of those four, training, evaluation and tool use on the most advanced models stopped.
This week on 𝐓𝐡𝐞 𝐀𝐈 𝐁𝐫𝐞𝐚𝐤𝐝𝐨𝐰𝐧:
𝐖𝐡𝐚𝐭 𝐡𝐚𝐩𝐩𝐞𝐧𝐬 𝐰𝐡𝐞𝐧 𝐚𝐧 𝐒𝐌𝐁 𝐝𝐨𝐞𝐬𝐧'𝐭 𝐣𝐮𝐬𝐭 𝐮𝐬𝐞 𝐀𝐈, 𝐛𝐮𝐭 𝐬𝐭𝐚𝐫𝐭𝐬 𝐞𝐦𝐩𝐥𝐨𝐲𝐢𝐧𝐠 𝐢𝐭?
Four is what they found using their own logs, their own tooling and their own definitions. Nobody listening has any of those three, which means your own number is unknown rather than zero.
Inside the episode:
🔴 A model computing a nine loop amplitude in theoretical physics, and what that ceiling means for the hardest job in your business
🔴 A serving system pushing more than a billion tokens a minute through a single chip, at under 6 cents per billion
🔴 An ultrafast mode producing roughly 750 output tokens a second, about 14 times standard inference
🔴 The three word test for observability: what an agent saw, what it decided, what it changed
🔴 Why 13,000 words of analysis landed on permissions rather than capability
Capability went up on three axes. Oversight went up on one. 🎙️
Stream on :
▶️ Youtube: https://www.youtube.com/@SMBAIHUB
▶️ Spotify: https://open.spotify.com/show/03455lXXM29joyuucQkkyq
▶️ Amazon Music: https://music.amazon.in/podcasts/b93fd52f-bc53-4ee9-98d7-3b9e10ea0e6f/ai-breakdown
▶️ Apple Podcast: https://podcasts.apple.com/in/podcast/ai-breakdown/id6799578599
▶️ iHeart Radio : https://www.iheart.com/podcast/1333-ai-breakdown-340815159/
▶️ Web: smbaimagazine.com
▶️ Web: smbaipodcast.com
Disclaimer: The information shared in this podcast is for general informational purposes only and should not be considered as direct financial or business advice. Always consult with a qualified professional for advice specific to your situation. This is an Independent commentary. Third-party excerpts are credited on screen or in the show notes. Some voices in this program are AI-generated.