This episode explores whether we're entering the era of vertical AI models, featuring major developments from Anthropic, Google, OpenAI, and specialized AI companies. The discussion centers on leaked details about Anthropic's Claude Mythos model and significant announcements from companies like Intercom and Cursor about their custom AI models outperforming general-purpose systems.

Key developments include Anthropic's accidentally leaked Claude Mythos described as their most powerful model ever, Google's real-time voice model Gemini 3.1 Flash Live, and OpenAI's decision to shelve adult content features. The episode examines how companies like Intercom and Cursor are achieving superior performance with domain-specific models built on open-source foundations, challenging the dominance of general-purpose frontier models.

The analysis draws heavily on The Bitter Lesson by Rich Sutton to understand how computational approaches consistently outperform human-knowledge-based systems, while exploring whether experience-based post-training might represent a new phase in this fundamental AI principle.

Anthropic's Claude Mythos Leak Reveals Most Powerful Model

A data leak revealed Anthropic is testing Claude Mythos, described in an unsecured blog post as 'by far the most powerful AI model we've ever developed'

Mythos represents 'a new tier of model, larger and more intelligent than our Opus models' with dramatically higher scores on software coding, academic reasoning, and cybersecurity tests

The model is 'very expensive for us to serve and will be very expensive for our customers to use' - Anthropic is working to improve efficiency before general release

Close to 3,000 unpublished Anthropic blog assets were found in the unsecured database, with the model also referred to by the codename 'Capybara'

Google's Real-Time Voice Model and OpenAI Product Decisions

Google launched Gemini 3.1 Flash Live, bringing real-time dialogue to voice models with continuous back-and-forth rather than jarring turn-based interactions

The model shows step change improvement on audio benchmarks including multi-step function calling, with Home Depot already deploying it for handling complex alphanumeric product codes

OpenAI shelved adult mode indefinitely after their advisory council unanimously opposed it due to 12% age detection failure rate and concerns about emotional dependence

OpenAI upgraded Codex with plugins while taking advantage of Claude's usage limit changes, resetting 'Codex usage limits across all plans to let everyone experiment'

The Bitter Lesson and Vertical Model Performance

The Bitter Lesson by Rich Sutton argues that 'general methods that leverage computation are ultimately the most effective' - brute force approaches consistently beat human-knowledge-based systems across chess, Go, and language

Bloomberg GPT, a 50 billion parameter finance-specific model, 'got absolutely smoked by the general models' despite being purpose-built with proprietary financial data

The question emerged whether last-mile usage data from companies like Cursor could change this equation through post-training rather than pre-training from scratch

Cursor's Composer 2 model beat Opus 4.6 on coding benchmarks while being much cheaper, built from Kimikate 2.5 with reinforcement learning representing 'three-quarters of the compute spent on the final model'

Intercom's Apex Model Beats Frontier Models

Intercom's CEO called their Finn model 'objectively the highest performing, fastest, and cheapest model for customer service, beating the very best models in the industry, including GPT 5.4 and Opus 4.5'

The Apex model achieved 2.8% higher resolution rate, 65% reduction in hallucinations, dramatically faster performance, and significantly lower costs than competing models

Success was attributed to 'domain-specific proprietary evals from our billions of human and agent customer service interaction data points' creating a continuous improvement flywheel

Decagon reported that 'over 80% of model traffic at Decagon now runs on models we've trained in-house, structured as a network of specialized models handling different parts of the interaction'

Industry Shift Toward Vertical Specialization

Andre Karpathy predicted 'more speciation in the intelligences' noting 'you don't need this oracle that knows everything. You kind of speciate it, and then you put it on a specific task'

Companies including Pinterest, Airbnb, Notion, and Cursor are 'finding it better, cheaper, faster to use and train open models themselves rather than use APIs for many tasks'

The trend suggests 'classic disruption is now at their door' for frontier labs, with the only way out being to 'disrupt themselves by building cheaper specialized models too'

Rich Sutton himself discussed this evolution on the Dwarkesh podcast, predicting systems that 'learn from experience' could 'perform much, much better and be much more scalable' than human-knowledge-based approaches