#252 - GPT 5.6, Grok 4.5, Nemotron-Labs-Diffusion, AI 2040
Last Week in AI
22 HOURS AGO
#252 - GPT 5.6, Grok 4.5, Nemotron-Labs-Diffusion, AI 2040
#252 - GPT 5.6, Grok 4.5, Nemotron-Labs-Diffusion, AI 2040
Last Week in AI
22 HOURS AGO
This episode covers a busy week in AI, featuring the rollout of GPT-5.6, new competitive models from SpaceXAI and Meta, and significant policy and safety discussions. The hosts analyze the implications of these developments, including pricing wars, regulatory inconsistencies, and emerging risks from Chinese open-source models.
The hosts discuss OpenAI's release of GPT-5.6 and its rebranded 'ChatGPT Work' tool, criticizing the US government's ad hoc regulation that allowed this launch while delaying Anthropic's Fable. They note that UK safety tests found GPT-5.6 easily jailbroken. The episode covers a price war, with SpaceXAI's Grok 4.5 and Meta's Muse Spark 1.1 offering aggressive pricing and strong coding benchmarks, though safety documentation is minimal. Meta's Muse Image and Video faced backlash for generating images of Instagram users, leading to a quick rollback. Chinese open-source models now command over 30% of OpenRouter tokens, but the hosts warn this may be temporary as Beijing considers restricting overseas access. Other topics include Meta's potential cloud business, energy grid strain from data centers, and Anthropic's new 'global workspace' interpretability method. The episode concludes with the AI 2040 proposal for US-China cooperation to slow AI development until alignment improves.
00:10
00:10
Jeremie will cover more details on recent AI news
01:34
01:34
The episode covers GPT-5.6, Grok 4.5, Muse Spark 1.1
08:12
08:12
Undisclosed internal deployments pose the same risks as external ones.
18:48
18:48
Token prices are dropping significantly
23:12
23:12
Meta's Cybench score jumped from 65% to 93%.
27:22
27:22
One model is cheaper than Grok 4.5
34:41
34:41
Safety institutes are crucial for detecting sleeper agents.
43:58
43:58
Intense competition in cloud computing for AI workloads
46:35
46:35
Increased compute supply could drive down token costs
54:12
54:12
6x faster per forward pass than comparable models
58:00
58:00
Pre-training sets priors before dual-track activation.
1:07:35
1:07:35
Model forced thinking about suppressed concepts
1:09:30
1:09:30
Middle layers form a shared reasoning space
1:12:58
1:12:58
China may end open-source AI releases.
