S11E06 – Shelved AI Model and the RAM Price Squeeze

Episode Summary

Mike followed up on the Artemis moon mission from the week before, then moved to technology news. He talked about a new stool-based test for cancers, how fast AI is being adopted, and how his own team now uses AI to build tools and review client accounts.

The main story was an Anthropic AI model called Mythos that the company chose not to release because it was too good at finding security holes. A caller asked about OneDrive and cloud data usage, and Mike finished with a long warning about RAM and storage prices rising on everything with a chip, plus a tease about saving 40% on electricity with a smart thermostat.

Topics Covered

Medical tech and AI adoption

Mike described a new stool-sample test that uses machine learning to detect cancers, including melanoma recurrence risk, and said it is about 90% accurate compared with a colonoscopy and not yet widely available. He said a Stanford AI index shows AI software-engineering scores rising from a maximum near 60% to near 100% in under five years. He also said AI is the fastest-adopted technology in human history, and that he avoids social media and thinks doomscrolling is bad for people.

How Mike uses AI at work

Mike said his team built a system where one AI role coordinates, another builds a tool, and a code reviewer checks it before he sees it, and that results are now more than 95% accurate. He said an internal AI reviewed 150 Microsoft mailboxes in four minutes for red flags, a job that would otherwise take half a day. His advice for workers was to use AI tools where the employer allows them, since falling behind is the bigger risk, and to understand their limits, noting that AIs cannot count reliably or do math.

The Mythos model

Mike said Anthropic found its Mythos model so good at finding software flaws that it found thousands of previously unknown vulnerabilities in four days, in Debian and other operating systems and web apps. Anthropic shelved it, and under Project Glasswing gave about 40 organizations plus the government access to find their own issues. He said Anthropic released a dumbed-down version as 4.7, which he has used for about a week and found very good. He noted that the AI race means someone else may build something similar.

RAM and storage prices

Mike said RAM that cost $60 now costs $300, and that 24 sticks of server RAM he needs went from $25 to $90 each. He said that RAM is expected to stay expensive until about 2030, because new production plants take that long. He blamed demand from AI companies and a large promised RAM commitment from OpenAI. He predicts cheap, underpowered computers, such as machines with 8 GB of RAM when 16 GB should be the minimum, and compared it to the Windows Vista era. He also said a 1 TB solid-state upgrade went from $100 to $300 and that the next iPhone is expected to cost almost twice as much.

Smart thermostat savings

At the end, Mike said he has been building a smart home and that adding rules to a smart thermostat cut his electricity bill by 40%. He said he will go into this next week.

Calls

• Matt asked how to tell if his Microsoft 365 files are saved both to the cloud and to his hard drive. Mike said OneDrive saves to the cloud first and syncs to the PC, and that the status icons beside each file show what is local. To keep a full local copy for offline use, right-click the OneDrive folder in File Explorer and choose “always keep on this device.”
• Matt also asked how to limit iCloud backup traffic that pushed his cable data use over the cap. Mike suggested switching to a provider without a data cap, such as Starlink, choosing which folders sync to iCloud, finding what keeps modifying a large file so it re-uploads, and setting the computer’s data limit and reset date so it holds back automatic downloads.