Daily Digest — 2026-09-24

The speaker introduces Jev, a novel AI model developed by TypeSafe AI, a company founded by Diego Almeida, a former OpenAI researcher who contributed to instruction-following research for ChatGPT. Jev's core innovation lies in its divergence from generative AI models; it functions exclusively as a general-purpose classifier or decision-making model rather than generating text.

To illustrate Jev's utility, the speaker provides the example of scanning an email inbox with thousands of emails. A typical approach with an LLM (Large Language Model) would involve sending each email to the LLM, asking it to classify the email (e.g., spam, receipt, newsletter, urgent, needs response), and then receiving a generated text output, possibly in JSON format, along with an explanation. Most of this generated text, including the explanation, would then be discarded, as the primary need is often just a decision or classification.

Jev streamlines this process by skipping the text generation step entirely. When given an email, a user or system asks Jev 'typed questions'—effectively seeking direct classifications or probabilities. For instance, it can return a choice indicating the email's category, a score for its urgency, or the probability that a specific 'yes' or 'no' statement about the email content is true. Furthermore, Jev is designed to evaluate several such judgments in parallel, meaning a single email can be simultaneously checked for spam, urgency, response requirement, and category without generating a single paragraph of output.

This non-generative approach offers significant efficiency and cost benefits. TypeSafe AI currently prices Jev based only on input tokens, at an approximate rate of 4 cents per million tokens. This dramatic reduction in cost, compared to LLMs that charge for both input and output tokens, makes AI-powered decision-making dramatically cheaper. The speaker links this to the Jevons Paradox, where something becoming dramatically cheaper often leads to a massive increase in its consumption. The thesis is that by making AI judgment cheap, developers will start integrating it into many more parts of their software.

Jev itself is a closed, hosted model. TypeSafe AI claims it utilizes a new architecture involving a parallel sampler and a training method they call RLCD. However, specific details regarding its weights, parameter counts, and low-level architecture have not been publicly disclosed. The speaker anticipates that this non-generative, decision-focused AI model will likely inspire similar models from other companies, as the core idea of a reusable AI decision primitive is compelling. The video concludes by emphasizing that AI models don't always need to 'talk' or generate extensive text; sometimes, software simply needs a precise, efficient decision, which Jev aims to provide.

The video introduces an alternative to conventional AI prompting, drawing inspiration from Andrej Karpathy's perspective on effective AI interaction. The core idea is to move beyond simple, one-off prompts and instead build a more robust, systematic interaction model with AI tools like Claude. This approach is structured into three distinct layers, each designed to improve the AI's understanding, execution, and reliability.

Layer 1: The Spec

The first layer emphasizes creating a detailed specification rather than merely assigning a task. Traditional prompting often involves giving the AI a task and hoping it fully comprehends the intent. The proposed method suggests a collaborative process where the user works with Claude to define a comprehensive spec. This involves having the AI 'interview' the user to clarify objectives, desired outcomes, and any constraints or background information relevant to the task. By breaking the work into smaller, manageable checkpoints before execution, the AI gains a deeper understanding, which leads to better questions, more accurate results, and a clearer execution plan. This structured approach helps prevent misunderstandings and aligns the AI's output more closely with the user's expectations.

Layer 2: The Verifier

The second layer focuses on verification and defining what constitutes a successful outcome. Before the AI begins any substantive work, the user needs to explicitly define what a 'good result' looks like. This includes establishing clear metrics and criteria for success. Furthermore, the AI should be equipped with methods to check its own work. This can involve leveraging other AI models for validation, comparing outputs against real-world data, or executing predefined tests. By integrating self-verification mechanisms, the AI can independently assess its performance, identify potential errors, and self-correct, thereby improving the reliability and quality of its output. This layer transforms the interaction from a 'fire and forget' prompt to a more accountable and self-improving process.

Layer 3: The Environment

The final layer centers on establishing a persistent and knowledgeable environment for the AI. Instead of starting from scratch with every interaction, the AI should operate within a pre-configured workspace. This workspace would pre-load all necessary instructions, skills, and rules that Claude needs to operate effectively. This means that every time the user engages with the AI, it already possesses a foundational understanding of the user's preferences, ongoing projects, and established operational guidelines. This 'knowledge base' and skill set accumulate over time, making the AI progressively better and more efficient with each interaction. The goal is to move beyond simply writing better prompts to building an entire system around Claude that continuously learns and improves, leading to a more seamless and productive partnership between the user and the AI.

Earlier this month, Automattic, the company that owns popular platforms like WordPress.com, Tumblr, and WooCommerce, experienced a significant internal power struggle. Matt Mullenweg, the founder and co-creator of WordPress, was reportedly forced into a leave of absence by the company's Board of Directors. This move came at a time when Mullenweg was allegedly at the Burning Man festival, micro-dosing homelessness. The incident drew parallels to the recent firing of Sam Altman from OpenAI, where a founder was unexpectedly removed from their leadership position.

Upon Mullenweg's return, he received a resolution stating that within 50 minutes, the board would vote to remove him as CEO. His request for a few hours to have the proposal reviewed by independent legal counsel was denied. The board members involved in this attempted ouster included CFO Mark Davies, Board Member Sue Decker, retired four-star general Ann Dunwoody, and Toni Schneider, who was Automattic's first CEO and whom Mullenweg once called his 'business soulmate.' Davies was voted in as the interim CEO, and the company issued a statement expressing full confidence in his leadership.

Adding to the controversy, Automattic's valuation by BlackRock had reportedly declined by 83%, from $7.5 billion to $1.26 billion, over the past two years. This decline was linked to Mullenweg's legal battle with WP Engine over WordPress trademarks. WP Engine had accused Automattic of destroying evidence, claiming 20 months of Mullenweg's Signal history and two of his phones went missing. Mullenweg's internal communications suggested he believed the board's actions were related to these legal issues, specifically a theory that removing him might lead to a more favorable outcome in the lawsuit.

Mullenweg's response to the board's actions was swift and aggressive. He used internal communication channels (Slack) to accuse Davies and the board of conspiring against him. He also posted on social media, implying that the board members were attempting to 'destroy his life' and that Silver Lake, a private equity firm behind WP Engine, was involved. Crucially, Mullenweg highlighted that he possesses 84% of Automattic's voting power, meaning he could effectively replace the board members who voted him out. He then announced he would be moving some of his critical assets, likely referring to WordPress.org, a 501(c)(3) non-profit he personally owns and which handles WordPress plugin, theme, and security updates, off Automattic servers.

Within 33 hours, Mullenweg was back in control. He announced on Slack that the board was 'back in agreement' and that he was back as CEO. TechCrunch confirmed Mullenweg's return with 'full support of the board.' However, it was later revealed that the 'full support' meant that Mullenweg had fired the entire existing board, including Davies, Decker, Dunwoody, and Schneider. He also fired Davies (the interim CEO) and Andy Missan (Chief Legal Officer). During their brief period of authority, Davies and Missan had reportedly signed reciprocal severance agreements that would entitle them to 12 months of base salary, an accelerated vesting schedule for their equity, and a year of health coverage if fired. As a result, Automattic now owes these two executives over $8 million.

Automattic employees reported being kept in the dark about the events, with questions on Slack going unanswered and an internal anonymous forum being shut down. Mullenweg's blog post to employees, ostensibly a 'Major Life Announcement,' was about his intent to buy a houseboat, further fueling employee frustration and a sense of disrespect. This entire episode underscores the complexities of corporate governance in founder-led tech companies, particularly when founders retain significant voting control, creating a dynamic where traditional board oversight can be effectively overridden.

The conversation begins by noting that war greatly accelerates progress in various fields, citing World War II as a prime example, particularly in medical advancements. A significant breakthrough discussed is the discovery of penicillin. While Alexander Fleming is credited with its discovery in 1928, observing mold killing bacteria, he abandoned further research. The real therapeutic potential was developed by Howard Florey and Ernst Chain at Oxford. Florey even flew penicillin spores across the Atlantic in his jacket for wartime precautions.

However, it was American ingenuity that solved the critical challenge of mass-producing penicillin. Florey's team took weeks to produce enough for a single mouse, whereas American companies rapidly scaled production. This mass production was vital for the war effort and, subsequently, for the civilian population, paving the way for other antibiotics and stimulating further pharmaceutical research.

The discussion then shifts to chlorpromazine, the first antipsychotic. This drug was synthesized by the French company Rhône-Poulenc in 1950, initially as an antihistamine. The company was looking for new markets for their antihistamine compounds, which were already known for causing drowsiness as a side effect.

The accidental path to chlorpromazine's psychiatric use involved a French naval surgeon, Henri Laborit, a Dunkirk survivor who was experimenting with antihistamines to reduce surgical shock. In 1952, Laborit observed that chlorpromazine strikingly calmed patients without sedating them prior to surgery. Recognizing this unique calming effect, he urged psychiatrists to try it on the mentally ill.

Psychiatrists Jean Delay and Pierre Deniker at Sainte-Anne Hospital in Paris, the largest psychiatric facility in the city, began testing chlorpromazine on psychotic patients on March 24, 1952. They quickly published six papers in six months, coining the term "neuroleptic" to describe its effects. This marked a paradigm shift in psychiatry, as chlorpromazine was the first drug to treat psychotic symptoms, not just agitation.

Initially, drug companies, including two before SmithKline & French, were hesitant to license chlorpromazine due to skepticism about a market for psychiatric drugs. However, SmithKline & French eventually licensed it in 1953. FDA-approved in 1954, it became known as Largactil in Europe and Thorazine in the US. The drug's rapid adoption was astounding; by 1964, over 50 million people worldwide had taken Thorazine. This success transformed SmithKline & French from a modest drugmaker into a pharmaceutical giant, now part of GSK. Thorazine alone earned $10 million in its first year and accounted for 18% of SKF's sales by 1960.

The patent for Thorazine expired around 1970, leading to struggles for SKF until Tagamet revived it. The success of chlorpromazine also spurred the development of minor tranquilizers like Miltown (promoted by Milton Berle as "Uncle Miltown"), and later Valium and Librium in the 1960s.

The speaker clarifies the terms positive and negative symptoms in psychiatry. Positive symptoms refer to experiences added to perception, such as hallucinations and delusions, and are often visible. Antipsychotics primarily help with these positive symptoms, either by dulling them down or reducing their intensity. Negative symptoms, on the other hand, involve a loss of abilities, such as apathy, lack of motivation, emotional expression, social drive, and poverty of language. Antipsychotics remain largely ineffective against these debilitating negative symptoms.