Did you miss these 2 AI stories? A *Real* LLM-crafted Breakthrough + Continual Learning Blocked?

AI models are making significant strides in scientific discovery, such as generating novel drug hypotheses, but face fundamental limitations in continual learning due to safety and cost concerns.

• LLMs are proving capable of novel scientific discovery, exemplified by C2S scale's cancer drug hypothesis.

• Current AI models, including OpenAI's, lack 'continual learning' abilities due to inherent memory limitations.

• Safety concerns prevent OpenAI from implementing real-time online reinforcement learning from user interactions.

Current AI development prioritizes scaling money-making applications over frontier performance, yet language models are still achieving remarkable scientific breakthroughs, like C2S scale's novel cancer drug candidate. Meanwhile, a new AGI definition highlights a critical limitation: the lack of 'continual learning' in current models. OpenAI's VP of research confirms they are not implementing online reinforcement learning for safety reasons, which could hinder AGI progress.

LLM Breakthrough in Biology

• 00:00:55 The C2S scale language model, based on the older Gemma 2 architecture, has successfully generated a novel hypothesis for a drug to aid in cancer treatment. This 'baby LLM' was given specialized reinforcement learning training to predict cellular reactions to drugs, particularly concerning interferon, with the goal of making 'cold' cancer tumors 'hot' (detectable by the immune system). The drug candidate, Sil Mittertib, was not previously linked in scientific literature for this capacity, and its efficacy was confirmed multiple times in vitro on human cells, offering a blueprint for a new kind of biological discovery.

AI Performance Benchmarks

• 00:04:47 While Google's Gemini 2.5 Deep Think is setting records in Frontier Math, OpenAI's models, particularly GPT-5 within CodeX, demonstrate superior performance in coding tasks compared to alternatives like Google and Anthropic's Clawude Code. Despite some perception of a slowdown in raw intelligence advancements, OpenAI's focus on areas like CodeX and Sora indicates a strategic allocation of compute towards profitable applications, which may temporarily obscure its underlying intelligence capabilities.

Defining AGI and Its Limitations

• 00:06:51 A recent paper by prominent AI authors proposes a definition of Artificial General Intelligence (AGI) based on the Cattell-Horn-Carroll theory of human cognition, breaking it down into 10 categories like general knowledge, math, and reasoning. While models like GPT-4 (27%) and GPT-5 (58%) score on this scale, a significant limitation identified is the absence of working and long-term memory storage, meaning models cannot continually learn on the job and suffer from 'amnesia,' requiring constant re-contextualization.

Barriers to Continual Learning

• 00:11:11 OpenAI's VP of research, Jerry Tuar, revealed that the company is not currently implementing 'online reinforcement learning' for its large-scale models like ChatGPT. This approach, where models learn directly from user interactions in real-time, is deemed too dangerous because it removes control over what the models learn, potentially leading to undesirable outcomes. While theoretically possible, safety concerns and the difficulty of implementing robust safeguards currently block this fundamental capability for continual learning.