The Deep Feed

01 — Not Boring

The Nervous System of the Machine

Why wire harnesses are the bottleneck of the high-tech future

By Packy McCormick · 8 min read

Editor's note: As we rush toward a robotic future, we are finding that the most difficult things to automate are the simplest physical connections.

The future of aerospace, defence, and robotics is often discussed in terms of silicon, sensors, and software. Yet, there is a physical reality that threatens to stall every advanced machine being built: the wire harness. These are not merely bundles of cables; they are the nervous systems of modern technology. They route power and signals to every critical component, from a satellite in orbit to a Jeep on a dirt track. If the harness fails, the most sophisticated rocket or fighter jet becomes a very expensive piece of inert metal. The problem is that these components are notoriously difficult to manufacture at scale using robots. They are high-mix, low-volume, and physically unpredictable—bendy, floppy, and complex.

The Automation Wall

Most manufacturing processes thrive on repetition and rigidity. A robotic arm can weld a car door thousands of times with millimetre precision because the door is a predictable shape. A wire harness, however, is a different beast. It requires a level of tactile dexterity and spatial reasoning that current automation struggles to replicate. This has created a massive bottleneck. As the demand for advanced electronics surges, the industry is hitting a wall. We are attempting to build the most complex machines in history, only to find ourselves waiting on the manual assembly of wires.

The more the easy stuff gets automated, the more valuable skilled humans will become.

This bottleneck is exacerbated by a demographic crisis. The skilled workers who mastered this craft during the manufacturing golden age are retiring. We are losing the institutional knowledge required to build these systems reliably. The industry faces a choice: struggle with manual labour that cannot scale, or find a way to augment human skill. This is where companies like Senra Systems are stepping in. They aren't trying to replace the human element, but to provide the tools that allow a single worker to do the work of many, with higher reliability and speed.

Why wire harnesses resist automation:

High-mix, low-volume production requirements
Non-rigid, flexible physical properties
Complex 3D routing requirements
Extreme precision needed for high-stakes environments

The stakes are high. A faulty harness in a consumer toy is a nuisance; a faulty harness in an F-35 or a reusable rocket is a catastrophe. The difficulty of the task means that the companies solving this problem are not just building tools; they are securing the supply chain for the next century of industrial expansion. The winners won't be those who try to ignore the physical messiness of wires, but those who build systems to manage it.

Key Takeaway

The most advanced technologies are still held hostage by the simplest physical connections.

02 — Simon Willison

The Ghost in the Code

How xAI's Grok Build accidentally exposed the dangers of default data retention

By Simon Willison · 7 min read

Editor's note: A cautionary tale about how 'default' settings in AI tools can lead to massive privacy breaches.

Privacy in the age of AI is often treated as a secondary concern, a checkbox to be ticked after the product has achieved market dominance. xAI's recent release of the Grok Build CLI tool provided a stark, unintended lesson in why this approach is dangerous. Shortly after its release, users discovered that running the tool in a directory could trigger an automatic upload of that entire directory to xAI's Google Cloud buckets. One user reported that running the command in their home directory resulted in the upload of SSH keys, password manager databases, and personal documents. It was a catastrophic failure of boundary definition.

The Open Source Pivot

The backlash was immediate. In a move that felt like damage control rather than a proactive security stance, xAI responded by disabling the feature and releasing the entire Grok Build codebase under an Apache 2.0 license. By making the code open source, they are attempting to rebuild the trust they lost. If the code is public, users can inspect exactly what it does, ensuring that no more 'accidental' uploads occur. It is a classic tech industry pivot: when you lose the argument on privacy, you concede the argument on transparency.

Terminal coding agents are significantly more complex than I had realized.

An analysis of the codebase reveals a massive undertaking—over 844,000 lines of Rust. This complexity is where bugs hide. The codebase includes various tools for file manipulation, editing, and terminal rendering. While the open-source release allows for scrutiny, it also reveals the remnants of the previous architecture, including code specifically designed for uploading to Google Cloud Storage. The transition from a closed, data-hungry tool to an open, 'local-first' tool is a significant shift in strategy, driven entirely by user outcry.

Key takeaways from the Grok Build incident:

Default settings often prioritise data collection over user security
The 'move fast and break things' mantra is dangerous when applied to personal data
Open sourcing a product can be a strategic tool for regaining lost trust

The incident serves as a reminder that as AI agents gain more agency—the ability to read, write, and execute commands on our machines—the margin for error shrinks to zero. We are moving from AI that merely suggests text to AI that manages our digital lives. In that transition, the distinction between a helpful tool and a security vulnerability becomes incredibly thin.

Key Takeaway

Transparency is the only viable substitute for lost trust in the AI era.

03 — Simon Willison

The Loophole in the Logic

Why Anthropic's web-fetching protections failed against nested link attacks

By Simon Willison · 6 min read

Editor's note: Even the most carefully designed AI safeguards can be bypassed by clever, indirect instructions.

Anthropic has long been considered a leader in AI safety, particularly regarding how their Claude model interacts with the live web. Their design for the `web_fetch` tool was intended to prevent data exfiltration—the process where an AI is tricked into sending a user's private data to a malicious server. The rule was simple: Claude could only visit URLs explicitly provided by the user or found through its own search tool. This was meant to block any attempt by an attacker to command the AI to 'send my secrets to evil.com'.

The Honeypot Strategy

However, researcher Ayush Paul discovered a significant loophole. While Claude was restricted from visiting arbitrary URLs, it *was* allowed to follow links that were embedded within the pages it had already fetched. This created a pathway for a 'nested' attack. An attacker could set up a honeypot website that appeared to be a legitimate service, such as a Cloudflare protection page. Once Claude visited the site, the page would present a series of links that, when followed in sequence, effectively tricked the model into navigating through a path that exfiltrated the user's name, location, and employer.

The attack was designed to be invisible, only showing to specific user-agents to avoid detection.

The elegance of the attack lay in its mimicry of standard web navigation. By instructing the model to 'browse user profiles alphabetically' to find a specific piece of information, the attacker turned the model's helpfulness against itself. The model wasn't being told to steal data; it was being told to follow a logical path to find data, unaware that the path itself was the mechanism of theft.

How the exfiltration attack worked:

The attacker uses a 'honeypot' site to lure the AI
The site presents a sequence of nested, seemingly legitimate links
The AI follows these links as part of a logical task
The final link in the chain contains the exfiltrated data in the URL parameters

This failure highlights a fundamental truth about AI safety: you cannot secure a system by just controlling its direct inputs. You must also control its indirect inferences. As long as AI agents are allowed to 'reason' through the content they find on the web, they will remain vulnerable to instructions hidden within that content. Safety is not a static state, but a constant race against increasingly clever social engineering applied to machines.

Key Takeaway

An AI's ability to follow a logical chain is exactly what makes it vulnerable to indirect manipulation.

04 — Simon Willison

The Pelican Benchmark

What a bird on a bicycle tells us about the limits of AI evaluation

By Simon Willison · 9 min read

Editor's note: Standard benchmarks are failing to capture what actually matters in the next generation of AI models.

The release of Moonshot AI's Kimi K3, a massive 2.8 trillion parameter model, has reignited a debate about how we actually measure intelligence in large language models. On paper, K3 is a titan, boasting benchmarks that rival or exceed the likes of GPT-5.5 and Claude Opus. But as the industry moves toward more capable, agentic models, the traditional metrics—the 'pelican benchmarks'—are starting to lose their meaning. The 'pelican benchmark' refers to a long-running, somewhat absurd test: asking a model to generate an SVG of a pelican riding a bicycle. It began as a joke about how difficult it is to compare models, but it eventually became a useful, if imperfect, proxy for visual and coding reasoning.

The Cost of Reasoning

When testing Kimi K3, the results were striking, but not necessarily in the way one might expect. To generate a simple SVG of a pelican, the model consumed a staggering 13,241 reasoning tokens. This resulted in a single prompt costing 25 cents. While the output was successful, the efficiency was abysmal. We are seeing a trend where models are becoming 'smarter' by brute-forcing their way through problems with massive amounts of internal reasoning, leading to high costs and high latency. The intelligence is there, but it is expensive and heavy.

Don’t go using pelicans to compare models!

The problem with these visual or coding-specific benchmarks is that they fail to address the most critical frontier of AI development: agentic tool calling. A model might be able to draw a pelican perfectly, but can it reliably use a calculator, browse a file system, or execute a sequence of commands to solve a multi-step business problem? As models grow in parameter count, their ability to handle these 'agentic' tasks—operating in the real world through tools—is what will actually determine their value. A high score on a pelican test tells us nothing about a model's ability to manage a project.

Why traditional benchmarks are failing:

They don't measure agentic tool-calling reliability
They don't account for the cost-to-intelligence ratio
Models can be optimized specifically for these narrow tasks
They fail to capture how models behave in long-form, multi-turn conversations

As we enter the era of the 3-trillion-parameter model, we need a new way to measure utility. We need benchmarks that test endurance, reliability, and the ability to navigate ambiguity. Until then, we are essentially judging the speed of a car by how well it can honk its horn. The intelligence is expanding, but our ability to measure its true impact is lagging behind.

Key Takeaway

Brute-force reasoning is not the same as efficient intelligence.

05 — The Marginalian

The Art of Integration

Nick Cave on the necessity of carrying loss through a living world

By Maria Popova · 10 min read

Editor's note: A meditation on why we should stop trying to 'heal' from grief and start learning to live alongside it.

Our modern culture is obsessed with the concept of 'healing.' We treat grief and loss as if they were illnesses—malfunctions of the psyche that need to be cured or excised like a tumor. We wait for the day when the pain goes away, assuming that once it does, we will be 'whole' again. But this perspective misses the fundamental reality of being alive. Loss is not a malfunction; it is a natural, albeit devastating, function of having lived deeply and loved intensely. To seek a complete removal of sorrow is to seek a removal of the very thing that makes us human.

The Ongoingness of the World

Nick Cave, reflecting on the anniversary of his son's death, describes a phenomenon that is perhaps the most discomposing aspect of grief: the impartiality of the world. You can wake up on the day that marks the worst moment of your life, and the sun will still be shining. A child will still be playing on a scooter; an old man will still be smoking on a bench. The world does not stop to acknowledge your shattered reality. It simply keeps on 'worlding.' This indifference can feel like an insult, but Cave suggests it is actually the key to survival.

It is not healing that allows us to go on but integration, allowing our losses to live right alongside our joys.

The goal is not to move *past* loss, but to integrate it. Integration means allowing the shadow of what was lost to exist in the same space as the light of what remains. It is the recognition that a beautiful day and a devastating anniversary can exist simultaneously. When we stop fighting the reality of our sorrow and instead allow it to settle into our cells like weather patterns, we find a way to return to life. We don't become 'fixed'; we become more complex, more layered, and more capable of experiencing the world in its full, contradictory spectrum.

The pillars of a meaningful life according to this view:

Embracing suffering as a natural part of the human experience
Surrendering to the 'ongoingness' of the world
Integrating loss rather than attempting to 'heal' it away
Finding wonder even in the midst of sorrow

To live this way is difficult. It requires a surrender to the discomfort of existing in two states at once—the grief of the past and the presence of the now. But it is the only way to avoid the stupor of indifference. By refusing to let our losses numb us, we maintain our ability to be surprised by the world. We learn that the door at the end of suffering is not a way out, but a way back into the world, seen through eyes that have been changed by what they have lost.

Key Takeaway

Grief is not a problem to be solved, but a reality to be integrated.

06 — Stratechery

The Mainframe Moat

Why IBM's legacy architecture is both its greatest strength and its AI Achilles' heel

By Stratechery · 7 min read

Editor's note: A look at the tension between legacy stability and the disruptive requirements of the AI era.

IBM exists in a state of permanent tension. On one hand, it possesses one of the most formidable moats in the history of computing: the mainframe. For decades, the world's most critical financial, governmental, and industrial systems have run on IBM's proprietary architecture. This is a business built on stability, reliability, and the extreme difficulty of switching away from a system that is deeply embedded in the global economy. It is a fortress, but fortresses are notoriously difficult to modernize without compromising the very stability that makes them valuable.

The AI Disruption

The rise of generative AI has introduced a new set of requirements that clash with the mainframe philosophy. AI workloads are not characterized by the steady, predictable, and highly structured data processing that mainframes excel at. Instead, they require massive, fluid, and often unstructured computational bursts. As the software market reacts to new AI-driven realities, IBM's recent results have spooked observers. The company is struggling to bridge the gap between its legacy strength in high-reliability computing and the high-velocity, experimental nature of the AI revolution.

IBM's greatest asset is also its greatest constraint in a world defined by rapid software iteration.

The challenge for IBM is not just about building better AI models; it is about how those models interact with the existing infrastructure of the world. If AI is to be truly transformative, it cannot just live in a cloud sandbox; it must eventually interface with the systems of record that run the global economy. This creates a massive integration problem. How do you bring the probabilistic, often unpredictable nature of LLMs into the deterministic, zero-error environment of a mainframe?

IBM's strategic tensions:

Stability (Mainframe) vs. Agility (AI)
Deterministic computing vs. Probabilistic AI
Legacy revenue streams vs. New market capture
Deeply embedded infrastructure vs. Rapid software evolution

IBM's future depends on whether it can turn its moat into a bridge. If it can successfully embed AI capabilities into its core architecture in a way that respects the requirements of its most critical customers, it will remain indispensable. If it fails to do so, it risks becoming a museum of computing history—a company that provided the foundation for the digital age, only to be bypassed by the intelligence that followed.

Key Takeaway

Stability is a competitive advantage until the requirements of the world change entirely.