Your signal. Your price.
Palantir CEO Alex Karp dismissed Wall Street fears that large language models could replicate his firm's enterprise software, criticizing AI rivals as unlikable and culturally incapable.
The hosts critique Anthropic's rollout of its powerful Mythos AI model as Fable 5, calling it a staged media cycle of fear and release they say works every six months.
Anthropic's Fable 5 launch triggered intense backlash over strict safeguards that blocked biomedical researchers, a 30-day data retention policy for enterprise messages, and silent degradation of outputs for AI development queries.
Anthropic's system card revealed it silently nerfed Fable 5 for frontier LLM development using prompt modification and steering vectors, breaking benchmark assumptions and making research failures indistinguishable from intentional degradation.
Tom Davidson steelmanned Anthropic's position, arguing silent nerfing is necessary to maintain a leading lab's lead during an intelligence explosion, as allowing competitors to use the model for R&D would prevent a critical safety pause.
Anthropic walked back the silent degradation policy within 24 hours, telling Wired it would make AI development safeguards visible after acknowledging it made the wrong trade-off, though experts like Dean Ball predict lasting broken trust.
OpenAI may cut token prices per a Wall Street Journal report, potentially starting a pricing war, while Sam Altman's Slack message hinted their next model isn't yet at Fable 5's level according to The Information.
Tomasz Tunguz explains applications now use model orchestration, using a high-cost model like Claude Fable to create skills, then running them locally to drastically cut token costs.
Anthropic's internal data shows its Mythos model improved completion rates for open-ended coding tasks from under 20% to over58%, but also caused a regression in trivial task performance from 100% to just over 80%.
Anthropic's research suggests AI is accelerating AI development, with Claude now writing most of Anthropic's code, and Mythos making better 'next-step' research decisions than human researchers 64% of the time in flawed scenarios.
Ben criticizes Claude's overly personified and anxious alignment, contrasting it with OpenAI's more detached approach, and fears Anthropic will lobotomize the public release of Mythos for safety.
A prompt engineering trick reveals AI image models' latent biases: asking to 'restore' a non-existent 'strange' photo generates grotesque imagery, while changing 'strange' to 'beautiful' generates sexualized content.
Anthropic launched Claude Fable 5, its first 'Mythos-class' model, which Nathaniel Whittemore describes as 'fairly undisputedly the best AI model we have ever been able to use'.
Fable 5 significantly outperformed competitors on key benchmarks. On Swebench Pro it scored 80.3% versus GPT-55's 58.6%, and it achieved a 29.3% on the new Frontier Code benchmark, more than double Opus 48's 13.4%.
API pricing for Fable 5 is set at $10 per million input tokens and $50 per million output tokens, double the cost of Opus but less than half the cost of the Mythos Preview within Project Glasswing.
Felix Ryeberg of Anthropic argued Fable 5 signals a shift from users giving AI 'tasks' to assigning 'responsibilities' or autonomous loops, such as having an agent monitor all crash reports instead of just fixing a single bug.
Jeffrey Cannell states current smaller local models lack the quality for coding agents compared to frontier models, and the scaling trajectory points to ever-larger models, making local high-performance compute a niche.
The initiative's strategy is hierarchical modeling, starting with proteins, then cells, then whole biological systems. Zuckerberg says you cannot understand cells without first understanding protein interactions.
At Mariana Minerals, all company data resides in a web-accessible, integrated data frame with minimal access controls, and they use LLMs to query and navigate the information repository.
Nathaniel Whittemore observes that Anthropic's messaging on Mythos public release is confusing, walking back a 'coming weeks' promise from the Opus 4.8 announcement.
Testers find Mythos powerful but eye-wateringly expensive, running through millions of dollars worth of tokens quickly, with Anthropic currently subsidizing the cost.
Nathaniel Whittemore summarizes OpenAI's thesis: knowledge work suffers from 'strange abundance' where producing artifacts is cheap but finding context and coordinating information is costly, citing a McKinsey study that workers spend over 25% of their week on email.
OpenAI reports 72% of Coda knowledge workers produce weekly artifacts, with common non-coding tasks being research (41%), data analysis (27%), and business workflow implementation (15%).
OpenAI observes a shift from sequential to parallel task execution in Coda, with 50% of users now running multiple tasks simultaneously, enabling a single worker to orchestrate work streams like a small team.
OpenAI's new Coda role-specific plugins bundle apps and skills for sales, analytics, and other functions, with six plugins including access to 62 apps and 110 skills.
Microsoft released seven new AI models at Build, with MAI Thinking 1 positioned between Sonnet 4-6 and Opus 4-6 ranges, using a 1 trillion parameter mixture-of-experts architecture.
Ali Bakhouch notes MAI Thinking 1 uses zero synthetic data, learning reasoning and tool use fully during post-training, which is a harder but more controlled approach.
Mustafa Suleyman claims Microsoft frontier tuning for McKinsey's tasks delivered higher win rates than GPT-5.5 at 10x lower cost, positioning cost optimization as Microsoft's core enterprise strategy.
Chris Summerfield notes current AI systems lack continual learning - the ability to update knowledge on the fly like biological brains. This is a core unsolved challenge in AI research.
AI models behave like humans primarily because they are trained in two stages. First, on massive human-generated text and image data, then further optimized for human preference through techniques like reinforcement learning from human feedback.