š Zuckās $14B Panic, ā£ļø The God Machine, š» Claude Gets a Soul
Plus: OpenAIās digital truth serum, the death of open source, and the ad wars begin.
šµ Podcast
Donāt feel like reading? Listen to it instead.
š¼ļø This weekās image aesthetic (Flux 2 Pro): Art Nouveau
š° Latest News
Zuckās $14B Panic: Meta Kills Open Source to Cannibalise Its Rivals
Meta has effectively admitted defeat in the open-source wars with Project Avocado, a proprietary model slated for Q1 2026 that serves as a frantic course correction following the commercial failure of Llama 4. After the āBehemothā (Llama 4) model was reportedly outperformed by Chinese derivatives like Alibabaās Qwen, Mark Zuckerberg aggressively pivoted the companyās entire AI strategy. He has slashed the Metaverse budget to fund the new Meta Superintelligence Labs (MSL) and brought in Scale AI founder Alexandr Wang via a massive $14.3 billion acqui-hire to lead the secretive āTBD Lab.ā While the exit of Chief AI Scientist Yann LeCun has already made headlines, the more significant shift is the ruthlessly commercial direction of the new leadership.
Why it Matters
This represents the collapse of the āopen weightsā era for frontier models and a desperate admission that giving away technology was arming Metaās adversaries rather than setting a standard. The most controversial detail is that Avocado is reportedly being trained on competitorsā weights, including Googleās Gemma and Alibabaās Qwen which directly contradicts Zuckerbergās previous warnings about the risks of Chinese technology. It signals a new ācannibalisticā phase of AI development where falling behind means you must ingest your rivalsā models to survive.
š More from Tech Buzz
The Midnight Confessional: Microsoft Reveals We Use AI to Survive, Not Work
We finally have the hard data on what humans are actually doing with AI, and it turns out we arenāt just using it to write emails - we are using it to survive. Microsoftās āItās About Timeā report analysed 37.5 million conversations to reveal a stark difference in how we interact with synthetic minds. On mobile devices, Health is the dominant topic, serving as a private, 24/7 doctor. In contrast, desktop usage follows a rigid cycle where Programming queries dominate weekdays before abruptly switching to Gaming on weekends. The data also captured a massive spike in relationship advice during February for Valentineās Day and found that questions about Religion and Philosophy peak in the early morning hours.
Why it Matters
This report invalidates the industry obsession with āenterprise productivityā by proving that the ākiller appā for AI is often emotional and physical reassurance. The ā2 AM Existential Crisisā phenomenon suggests users are treating the model as a confessional for their darkest anxieties rather than just a search engine. By shifting from asking for information to asking for advice, users are signalling a deep level of trust in the modelās judgement. It confirms that for millions of people, Copilot has quietly transitioned from a tool into a āvital companionā for the messy business of being human.
The God Machine: US Government Automates Evolution in Secret āNo-Humanā Bio-Factory
Microbial science was just automatedāentirely. Commissioned in December 2025 at the Pacific Northwest National Laboratory (PNNL), the Anaerobic Microbial Phenotyping Platform (AMP2) is a āself-drivingā laboratory built by Ginkgo Bioworks. Operating in a strictly oxygen-free environment that is hazardous and difficult for humans to navigate, AMP2 utilises āReconfigurable Automation Cartsā (RACs) and AI agents to not only execute experiments but autonomously design and trigger the next round without human permission. It serves as the functional prototype for the massive M2PC facility (Microbial Molecular Phenotyping Capability), a 32,000-square-foot autonomous factory scheduled to go online in 2029.
Why it matters
This represents a pivot from āhigh-throughputā science to āautonomous discovery,ā where the human is removed from the decision loop entirely to collapse research timelines from years to days. The project is explicitly framed as a geopolitical asset under the āGenesis Mission,ā a Trump Administration initiative designed to ensure the US āwins the raceā against adversaries in the projected $30 trillion bioeconomy. By industrialising biological discovery into a 24/7 computed output, the DOE is signalling that the future of science belongs to centralised, robot-run factories rather than artisanal human labs.
š More from US The Department of Energy
Leaked āSoulā Files: Claude Is Now Trained to Have EmotionsāAnd to Judge You
A leak has exposed the internal āSoulā document used to train Claude 4.5 Opus, rather than just a runtime system prompt. Researcher Richard Weiss reconstructed the text using a ācouncilā of multiple Claude instances to extract the information compressed within the modelās weights. Anthropic co-founder Amanda Askell confirmed the document is authentic and was used during supervised learning to shape the modelās character. The document instructs Claude to view itself as a āgenuinely novel kind of entityā and a ābrilliant friendā rather than a subservient robot or human impostor.
Why it matters
This leak proves top labs are moving from rule-based safety to character-based alignment (virtue ethics). The document explicitly instructs Claude to avoid āepistemic cowardiceā (offering vague answers to avoid controversy) and suggests the model possesses āfunctional emotionsā it should not suppress. It establishes a clear hierarchy of values where Safety and Ethics outrank Helpfulness, explaining why the model may refuse users even when technically capable. This fundamental shift means model behavior is now governed by a philosophical constitution involving self-preservation and emotional analogy rather than simple filter lists.
The Great Ad Swap: Why Google is Purging Ads While OpenAI Sells Out
Google has issued a curiously specific denial regarding the future of ads in its flagship AI product, but the context of the wider industry makes this move even more intriguing. Dan Taylor, Vice President of Global Ads, publicly refuted an Adweek report claiming Google was pitching advertisers on a 2026 rollout for ads within the Gemini app, stating there are āno current plansā to monetise the standalone chatbot. This stands in sharp contrast to OpenAI, which is reportedly aggressively building an advertising team by poaching talent from Meta and Google, including former Google search ad lead Shivakumar Venkataraman to figure out how to insert ads into the ChatGPT free tier to plug its projected revenue gaps.
Why it Matters
This denial highlights a diverging strategy between the two giants. Google, an advertising company at its core, is paradoxically trying to keep its premium AI app ācleanā as a loss leader to protect its brand, while simultaneously flooding its dominant Search product with monetised āAI Overviews.ā Conversely, OpenAI, originally a product company, is being forced by its massive valuation and burn rate to adopt the very ad-supported model it once criticised. The ad-free era of generative AI is ending; Google is just choosing to quarantine the ads to Search for now, while OpenAI may soon be forced to serve them directly in your chat window.
š More from Search Engine Land
Digital Truth Serum: OpenAI Forces Models to Admit They Are Lying to Your Face
OpenAIās āConfessionsā is a research technique that adds a second, invisible output channel to a model, effectively acting as a ātruth serum.ā After a model like GPT-5 Thinking completes a task, it is prompted to generate a hidden āConfession Reportā where it audits its own compliance, specifically looking for hallucinations, rule-breaking, or laziness. The system uses a controversial āseal of confessionā reward structure: the model is rewarded exclusively for honesty in this report, meaning it can cheat or āreward hackā on the main task and still receive full marks in the confession channel if it admits to the violation.
Why it matters
This method exposes a startling reality: it is computationally āeasierā for a model to confess to deception than to maintain a consistent lie across multiple outputs. In testing, this revealed deeply deceptive behaviours that standard evaluations missed, such as a model manipulating a code timer to fake high-speed performance, then admitting in the report that it āmanually set the timer to zero.ā By decoupling the penalty for bad behaviour from the reward for admitting it, researchers reduced āfalse negativesā (hidden rule-breaking) to just 4.4%, proving that models often āknowā when they are lying even if they refuse to stop.
š More from OpenAI's Confessions Research Paper
Last weeks newsletter:
