Caleb Writes Code — searchable YouTube transcripts & AI Q&A

Reinforcement Learning sucks..

Reinforcement learning remains more art than science, with key bottlenecks including the exploration-exploitation tradeoff, data scarcity, sample inefficiency, and the difficulty of designing reward systems that don't lead to reward hacking. The speaker argues that throwing more compute or data at RL doesn't guarantee better results, unlike in pre-training, and that progress depends on making RL more generalizable beyond narrow tasks. Despite these challenges, RL has driven significant improvements in LLMs, particularly in reasoning models.

MiMo V2.6 explained in 12min..

Xiaomi's MiMo V2.6 post-training cost $2.6 million for the Pro model and $854,000 for the Flash, taking under 6 days with thousands of GPUs. The key challenge is coordinating asynchronous reinforcement learning across GPUs, which they solve using partial rollout, policy staleness compensation, and a dynamic sampler that mixes successful and failed rollouts. The model generates billions of tokens per training step, and the Pro model broke the Pareto frontier in cost efficiency.

OpenAI solved hundreds of math problems..

The speaker argues that AI's rapid production of research, exemplified by OpenAI releasing 722 manuscripts in a single day, is overwhelming the traditional peer-review system and shifting the bottleneck from discovery to verification. He suggests that while AI can generate discoveries, humans must decide which are meaningful, and the current academic structure is ill-equipped for this change. He notes that some mathematicians see this as a golden era, while others fear for their careers, and that the tension may require using AI itself to help validate work.

Haiku is as cheap as GPT-6 Luna..

Haiku 5.5's two-tiered pricing makes it cost-effective for high-volume tasks under 100,000 tokens, but its price advantage diminishes at higher intelligence levels compared to models like GPT-6 Luna. The speaker demonstrates using Haiku 5.5 in his personal agent Attekus for tasks like pulling up research papers and changing Excel to dark mode, noting that while it handles simple commands well, it struggles with complex multi-step web navigation and can be slower than manual operation. He argues that such cheap, intelligent models are reshaping decisions about local hardware versus API usage.

Gemini 4 Argon explained in 5min..

Gemini 4 Argon's benchmark scores, especially on coding, are strong but may be inflated due to public benchmark contamination, the speaker argues. He notes that Google's pricing advantage is temporary and that the model falls below the Pareto frontier compared to Anthropic and OpenAI. The speaker also highlights Google's strategic focus on monetizing existing infrastructure rather than subsidizing subscriptions, and the significance of extending output tokens to 1 million for long-horizon tasks.