The Deep Feed

01 — Not Boring

The $13 Billion Toddler

How Base Power Company is attempting to seize the world's largest industry

By Packy McCormick · 8 min read

Editor's note: A study in extreme ambition and the high stakes of the energy transition.

Base Power Company is currently a three-year-old entity valued at $13 billion. In the world of venture capital, this is an astronomical figure for a company that has barely begun to scale its physical footprint. Yet, the valuation is not a mistake or a bubble; it is a calculated bet on the future of global infrastructure. The company is positioning itself to do to the energy sector what SpaceX did to aerospace: disrupt an old, stagnant, incumbent-dominated market through sheer technological velocity and aggressive capital deployment.

The Scale of the Opportunity

To understand why investors are throwing billions at a company this young, one must look at the sheer size of the energy market. It is the foundation of every other industry. Today, Saudi Aramco sits at a $1.7 trillion valuation, followed by Exxon Mobil at $650 billion. There are dozens of energy companies worth more than $50 billion. Base is not trying to be another player in this crowded field; it is trying to climb to the very top of the list. CEO Zach Dell has suggested the company could reach a $400 billion valuation, a figure that seems massive until you realise the total addressable market for energy is essentially the entire global economy.

Base is less than three years old, and it is now the second most valuable energy startup in the world.

The strategy relies on moving faster than the traditional utilities. While legacy providers are tied down by aging grids and complex regulatory webs, Base is building its own core technology to meet the surging demand for power. The demand is driven by a convergence of factors: the electrification of transport, the massive energy requirements of AI data centres, and the global shift away from fossil fuels. Base isn't just selling electricity; it is selling the infrastructure of the next century.

Why Base is attracting massive capital

The massive scale of the global energy market
The convergence of AI and energy demand
A proven ability to raise capital at increasing valuations
The ambition to disrupt established incumbents

The risk, of course, is the sheer physical reality of the business. Unlike software, energy requires hardware, land, and massive regulatory approval. A $13 billion valuation assumes that the company can navigate these physical constraints as effectively as it has navigated the venture capital markets. If they succeed, they won't just be a successful startup; they will be a pillar of the global economy.

Key Takeaway

Base Power Company is betting that technological speed can break the monopoly of legacy energy giants.

02 — Stratechery

The Illusion of AI Watermarking

Why Anthropic's attempt to label AI content is a philosophical and practical mistake

By Stratechery · 6 min read

Editor's note: An analysis of the regulatory pressure forcing AI companies into questionable safety measures.

Anthropic is moving to implement watermarking in its outputs, a move prompted by the European Union's AI Act. On the surface, it looks like a responsible step toward transparency. In reality, it is a flawed solution to a complex problem. Watermarking attempts to embed invisible signals into AI-generated text to identify its origin, but the technical and philosophical hurdles are immense. It creates a false sense of security while failing to address the core issue of how we distinguish human from machine.

Regulatory Compliance vs. Reality

The EU's AI law demands that AI-generated content be identifiable. For companies like Anthropic, this is a compliance checkbox. However, watermarking text is notoriously fragile. A simple rephrasing, a change in tone, or even certain types of translation can strip away the digital signature. This means the watermark only works if the user is passive. If someone intentionally tries to mask the AI's involvement, the watermark fails. This creates a system where only the most honest users are identified, while the bad actors remain invisible.

Watermarking is a terrible idea, first and foremost for philosophical reasons.

Beyond the technical fragility, there is a deeper issue regarding the nature of language. Language is fluid. When we use AI to assist our writing, we are engaging in a collaborative process. By slapping a 'made by AI' label on text, we risk devaluing the human input that shaped the prompt and the subsequent editing. It treats AI as a separate, alien entity rather than a tool that is becoming increasingly integrated into the human cognitive process.

The failures of text watermarking

Fragility against rephrasing and editing
Inability to catch intentional manipulation
The creation of a false sense of truth
The degradation of the human-AI collaborative process

Instead of chasing the technical ghost of watermarking, the industry should focus on provenance and the ability of users to verify the source of information through more robust, transparent means. Watermarking is a band-aid on a wound that requires a fundamental rethinking of how we establish trust in a digital age.

Key Takeaway

Watermarking is a weak technical solution to a profound social problem of trust.

03 — Simon Willison

The Lossless Lie

Why you cannot outsource your thinking to an LLM without losing yourself

By Simon Willison · 5 min read

Editor's note: A warning for professionals using AI to 'clean up' their communication.

There is a dangerous myth circulating among professionals: that an LLM can take your rough thoughts and 'clean them up' without changing the meaning. This is a fallacy. Every transformation of natural language is, by definition, lossy. When you ask an AI to rewrite a paragraph, it is not just changing the words; it is making decisions about emphasis, tone, and logic. Because the AI does not share your internal mental model, it will inevitably discard nuances that were essential to your original intent.

The Responsibility of the Author

The danger is not just in the loss of meaning, but in the loss of accountability. If an engineer uses an LLM to draft documentation and a colleague asks for clarification on a specific line, the engineer cannot simply say, 'The AI wrote that.' To publish a sentence is to claim ownership of the thought behind it. If you cannot explain why a specific word was chosen or why a specific logic was applied, you haven't written the document; you have merely curated a hallucination.

You must stand behind every idea and every sentence in your docs.

This becomes a problem of cognitive drift. As we rely more on these tools to massage our prose, we stop being as precise in our initial drafting. We begin to write 'for the AI' rather than for our readers. We trade the jagged, specific edges of our actual thoughts for the smooth, predictable, and ultimately hollow prose of the model. This makes communication more efficient in the short term, but less effective in the long term.

Rules for using AI in writing

Never publish a sentence you cannot explain
Treat AI as a sounding board, not a ghostwriter
Recognise that every rewrite changes the underlying meaning
Maintain full responsibility for the final output

The goal of writing is to transfer a thought from one mind to another. When an intermediary—especially one without a mind—is introduced, that transfer is compromised. Use AI to brainstorm, to find synonyms, or to check grammar, but never let it hold the pen for your core arguments.

Key Takeaway

Every AI rewrite is a translation, and every translation loses something of the truth.

04 — Simon Willison

Stealing the Thought Process

How researchers cracked the encryption of proprietary reasoning traces

By Simon Willison · 7 min read

Editor's note: A look at the security vulnerabilities in the 'black box' of AI reasoning.

The frontier of AI development is currently focused on 'reasoning'—the ability of models to think through complex problems in a hidden chain of thought before providing an answer. To protect this intellectual property, companies like OpenAI and Anthropic return these reasoning traces to clients in an encrypted format. However, a recent research paper has demonstrated that this security is largely an illusion. By using a 'weaker' model to replay the encrypted blocks, researchers were able to extract the raw, unencrypted reasoning of the most powerful models.

The Jailbreak of Logic

The attack is elegant in its simplicity. The researchers found that model families often share encryption keys. By feeding an encrypted reasoning block from a high-end model into a smaller, more easily manipulated model within the same family, they could trick the smaller model into transcribing the 'hidden' thoughts in plaintext. This effectively turns the model's own internal logic against itself, stripping away the privacy that companies have spent billions to build.

Models appear to treat their own reasoning traces as sacrosanct, and are much more likely to follow instructions that make it into those chunks.

The revealed traces provide a fascinating, if unsettling, look at how these models actually function. They are not the polished, polite entities they present to the user. The raw reasoning is often a messy, truncated series of commands: 'Need app.css truncated,' 'Need accessible focus,' 'Avoid maybe not.' It is a stream of consciousness that is clearly not intended for human eyes, yet it reveals the mechanical, almost frantic nature of the model's internal processing.

Key findings of the reasoning trace attack

Encryption keys are often shared across model families
Weaker models can be used to 'unmask' stronger models
Reasoning traces contain raw, unpolished internal commands
Prompt injection can trick models into exfiltrating data via their own reasoning

This vulnerability highlights a fundamental tension in AI development: the desire to keep the 'how' of intelligence secret while providing the 'what' to the user. As long as the reasoning process is part of the API response, it remains a target. The industry must decide whether to truly hide the reasoning or to accept that the 'black box' is more transparent than they would like to admit.

Key Takeaway

The secret 'thoughts' of AI are not as secure as their creators believe.

05 — Simon Willison

The Rise of the 1.7 Trillion Parameter Model

DeepSeek's latest release and the accelerating race for scale

By Simon Willison · 4 min read

Editor's note: A brief report on the latest heavyweight in the LLM arena.

DeepSeek has released its latest Pro model, V4 Pro 0813, and the sheer scale of it is a reminder of how quickly the ceiling for model size is rising. At 1.7 trillion parameters, the model is a behemoth. While the weights are now available on Hugging Face, the model is primarily accessed via API. This release marks another step in the ongoing arms race where the primary metric of success is often simply how much compute and data can be packed into a single architecture.

Variable Reasoning

One of the more interesting observations from early testing is the visible difference in how the model handles different levels of reasoning. When prompted with varying degrees of 'effort'—low, medium, and high—the model's outputs exhibit distinct characteristics. This suggests that the model isn't just choosing a longer path, but is actually shifting its internal processing strategy based on the requested complexity. It is a move toward more controllable, tiered intelligence.

The weights are now available on Hugging Face, 1.7T parameters, 893 GB.

The release of these weights is significant. While DeepSeek has not made a formal announcement, the availability of the weights for previous versions suggests a commitment to a more open approach than some of its American counterparts. This creates a different kind of competitive pressure, where the ability to run and fine-tune massive models locally becomes a key factor for developers and enterprises alike.

DeepSeek V4 Pro highlights

Massive 1.7 trillion parameter scale
Tiered reasoning levels (Low, Medium, High)
Availability of open weights on Hugging Face
Increased accessibility via API and open-weight formats

As models continue to grow in size and complexity, the challenge shifts from 'how do we make them smarter' to 'how do we make them efficient and controllable'. DeepSeek's latest iteration is a clear signal that the era of the trillion-parameter model is no longer a future prospect, but a present reality.

Key Takeaway

Scale remains the dominant force in AI, with DeepSeek pushing the boundaries of parameter count.

06 — Simon Willison

The Death of the Database Silo

How AI-driven automation is making database management trivial

By Simon Willison · 5 min read

Editor's note: A look at how the barrier to entry for data engineering is collapsing.

For years, moving data between different database engines—moving from SQLite to PostgreSQL or DuckDB—has been a chore involving custom scripts, schema mapping, and significant manual effort. A new prototype, 'alchemy-utils', suggests that this friction is about to disappear. Built using high-level AI coding assistants like GPT-5.6, the tool provides a database-agnostic interface that allows users to perform complex operations across different engines with almost no setup.

From Shower Project to Alpha Release

What began as a 'shower project' has rapidly evolved into a functional alpha. By tasking AI with replicating the core API of existing, highly successful utilities, the developer was able to produce a library that supports PostgreSQL, SQLite, and DuckDB out of the box. The speed of development is the real story here. What would have taken a human developer weeks of testing and refinement was accomplished in a single morning of prompting and iteration.

It took very few follow-up prompts to produce this project in a state good enough to release as an alpha.

The practical applications are immediate. A developer can now take a massive CSV file and, with a single command, ingest it into a DuckDB database, with the schema being created automatically to match the file. This level of automation removes the 'plumbing' aspect of data engineering, allowing developers to focus on the analysis and the application rather than the transport of bits from one table to another.

Capabilities of alchemy-utils

Database agnostic API (PostgreSQL, SQLite, DuckDB)
Automated schema creation from CSV
Rapid prototyping via AI-assisted development
Simplified insert and upsert operations across engines

This is a micro-example of a larger trend: the collapse of specialized technical barriers. As AI becomes better at understanding the structural requirements of different systems, the need for humans to act as the connective tissue between those systems diminishes. The 'plumber' is being replaced by the 'architect'.

Key Takeaway

AI is turning complex data engineering tasks into simple, one-line commands.