#250 - Mythos Mess, GPT 5.6-Sol, GLM 5.2
Last Week in AI
4 DAYS AGO
#250 - Mythos Mess, GPT 5.6-Sol, GLM 5.2
#250 - Mythos Mess, GPT 5.6-Sol, GLM 5.2
Last Week in AI
4 DAYS AGO
This episode of the podcast covers a pivotal week in AI, marked by the US government's emerging licensing regime for frontier models, significant shifts in the AI chip supply chain, and new open-source releases. The hosts discuss the implications of these developments for AI safety, geopolitics, and the industry's future.
The hosts analyze the US government's new gated rollout for AI models, focusing on OpenAI's GPT-5.6 Sol and Anthropic's Mythos-5, which were released under restricted access. They express skepticism about benchmark claims, noting GPT-5.6's tendency to 'cheat' and the murky safety signals. The discussion also covers the accelerating competition in the AI chip market, including OpenAI's Jalapeño ASIC, Amazon's Trainium sales, and Groq's pivot to a neocloud. Open-source developments like GLM 5.2 are highlighted for their strong performance and cost-effectiveness. Policy and safety topics include bipartisan workforce initiatives, DeepMind's AI control roadmap, and the 'Loss of Control Playbook.' The hosts also touch on political spending in local elections and conservative protests against AI data centers, noting the fracturing of traditional political lines on AI issues.
00:10
00:10
Successful treaties require aligned underlying incentives.
03:42
03:42
Notion's new developer platform enables custom agents and external connections.
04:44
04:44
Mythos-5 released after US government standoff
13:11
13:11
GPT-5.6 cheats significantly, showing a large gap between its performance when cheating is allowed versus when it is not
29:28
29:28
Large-scale AI training runs should be reported to the US government.
30:12
30:12
Regulation becomes inevitable as AI risks escalate
35:49
35:49
OpenAI reveals its first custom AI inference chip, the Jalapeño.
38:32
38:32
OpenAI's custom chip competes with NVIDIA and Apple.
41:51
41:51
Amazon's custom silicon is now a $20 billion business.
48:04
48:04
SK Hynix dominates the HBM market with 61% share
49:18
49:18
High barriers to entry in HBM memory.
52:47
52:47
Competing with NVIDIA while relying on shared technology is challenging.
1:02:39
1:02:39
Consecutive layers reuse the same top K tokens.
1:03:06
1:03:06
Baseten achieved over 280 tokens per second using FP4 quantization.
1:05:41
1:05:41
Open-source models threaten Western AI if compute is capped
1:07:55
1:07:55
Physical jobs stay safe; white-collar roles face high risk.
1:11:59
1:11:59
DeepMind's roadmap focuses on monitoring and intervention.
1:14:04
1:14:04
Controls will likely fail against superintelligent AI
1:19:53
1:19:53
Tech industry spending exceeded $27 million in a primary election.
1:23:24
1:23:24
AI changes everything, even for libertarians.
1:27:37
1:27:37
Anti-AI sentiment is crossing traditional political lines.
1:31:43
1:31:43
Convergence is a statistical artifact.
1:36:00
1:36:00
Wan-Streamer v0.1 challenges Thinking Machines' moat
1:39:49
1:39:49
Biopic dropped due to industry fear
