Issue 79 November 6th 2026

First Take

Synthesis

Last week, Nvidia released a new and very small large language model designed for non-commercial edge compute infrastructure. This would be the model Nemotron-3.5-bolt:e4b. Based on Google's Gemma 4 and quantized to between 4 to 6 bits based on criticality and fitting inside of 6GB of VRAM, this model can run within a percentage point of Gemma 4's intelligence even on my ancient mini-PC gaming rig at home while achieving throughputs up to 80 tok/s.

I'm reminded of how old game cartridge systems initially required vast memory storage, but shrank dramatically as engineers learned how to compress and swap memory on the fly. It was amazing what could be achieved with just a 16 bit system with only 512kb of active memory. Everything else was extended memory swapping and flash for program storage on the cartridges. Later, as the hardware pipelines became cheaper, big memory and big compute footprints resumed, but towards the end of each console processor type lifecycle, extreme efficiencies were again employed to provide the highest possible color gradients, video resolutions and framerates while delivering ever more interesting ways to interact with the games. Today of course, we have AR and VR interfaces for those with deep pockets, while large screens and keyboard, mouse or gaming pads serve the majority of us offering life-like renditions of whatever.

My point is that edge-based intelligence hardware requirements are shrinking at the same time that model capabilities are growing. This odd combination means I can now run a local agentic AI at home that could respond as quickly as most cloud models without shooting any inference compute into a cloud. I am as always, loathe to share my specific build, but the model choices alone should be information enough for those with real interest.

Still, I embarked on this effort with a few ideas about what I wanted the end product to be capable of. Extensible, fast, capable and useful for multiple IO types. After getting the LLM up and running, I extended its agentic capabilities to the extent the off the shelf open-source harness permitted with tightly controlled email, file and document access/edit/creation as well as web-search and browser interaction capabilities. image generation required another model for which I found Z-image-turbo at 512x512 about the best I could get to run at a decent speed while Nemotron was loaded. If I wanted load and unload models, I could scale up a bit, but that eats user interaction times.

Now, I found myself running two inference engines, two models and an extensible python framework for creating more tools. Exploring the capabilities, I found getting directions to be an issue. I tried to get Gemini on this and it attempted to build a custom agent skill for this work that failed miserably across multiple iterations and the last version that sort of worked took about 15 minutes to produce the not so useful directions. I knew there was another way to skin this cat, but had let Gemini do what it thought would work.

Giving up on Gemini, Claude helped me build a custom MCP server with Python that could translate a request for directions between two addresses into useful turn-by-turn directions. This relied on a couple of external Internet based resources that permit free lookups, but all inference work was local and it was FAST. A 52 mile, 28 turn set of directions took 15.4 seconds to produce. The actual Internet lookup portion was only about 5 seconds with the rest being inference to parse the request, convert it into an appropriate query of the two Internet resources being used and parsing the returned data to deliver the actual directions. Three actions: 1. Parse and query. 2. Lookup and retrieve. 3. Return and parse. There were also some options for finding addresses that weren't available in the public mapping resource by converting to lat/long coordinates via a separate lookup and a fall-back error mode to request cross streets or coordinates if either lookup failed. My test case included 1 address that wasn't searchable in the public database and had to be converted to lat/long. So, 15 seconds really was a kind of worst case time-frame.

Now you say, so what? Google maps or Waze can do that super fast with zero issues. You're right. They can, but the difference is that this tool doesn't necessarily have to even touch the Internet. Both of the sources being queried are publicly accessible databases that you can copy and host locally. You could essentially replicate the text portion of Google Maps turn-by turn direction feature on local hardware capable of interpreting your speech or text request for directions with plain language understanding. I know. There are already commercial software sources that can do that. You're right. There are. Keep in mind however, that what I did, didn't expose all my data to a commercial collector of it or fund any such.

Achieving that capability with cash or data collection was not the point. The point was to synthesize my understanding of AI systems to replicate an available capability with no token rentals or software costs. Not even the hidden costs where your data is the product. Some of you reading this will just say, "Meh, Claude did the heavy lifting in writing your MCP server". You'd be right. Using AI to aid in this doesn't detract from helping me understand it. I don't need to learn how to code perfect python when I can read it quite well vs. beating my head against syntax errors and missing/extra commas, brackets or braces for hours on end. I COULD have actually written this. Claude simply sped the process for me. I might have even asked the local model to try it, but was loathe to trust it with the task at this early stage. This was mostly due to the constrained context window of the model at 256k tokens.

I might test it later to see how it does comparatively, but I had already wasted enough time with Gemini's custom agent skill efforts, that I was kind of through with their models for this work. That would include Nvidia's highly customized and quantized Gemma 4 model that I had running locally. Supposedly, Nvidia had this model scoring quite well on coding skills. I will probably find out later. I mean, one of the ways to improve models is by using the model to improve itself, so at some point, I'll have to have the snake eating itself to see what we get.

If you learned nothing else from this article, you learned that I spend a great deal of my weekends playing with systems and tools for no good reason other than to learn about them and how to best use them. That is what I really want to share. A qualification standard if you will, to help you realize that what I write in The Shift Register is based on real experiences, research and tested outcomes. While I permit AI voices in this newsletter, that has always been just a unique feature that is more a guided exploration of model differences and growth than anything else. The meat and potatoes of my work here has always been the curated and self-written articles featuring my latest research and experience derived opinions and advice. I hope you, our readers can appreciate this level of work and find some use for it as the helpful insight I want it to be.

Kudos to Gemini for the graphic.

Editorial

CIO's Corner

With all the current news focus turning to edge-based compute for AI infrastructure as security and cost control measures, I thought it would be nice to go back to basics in terms of implementing AI in our organizations with intention, controls and proper auditing. That first part is what earns the value of AI vs. random adoption across functional groups, so let's start there.

Implementing AI with intention means you have identified specific use cases for AI implementations where you understand the potential costs and efficiency gains well enough to determine an accurate ROI over a given time frame. I know that sounds very difficult in a world where the latest AI model is obsolete in 6 months, but obsoletion does not impune older model support for a solution that is working. In fact, smaller models that are specifically tuned for your solution are going to be where you can experience the greatest efficiency gains. Just renting tokens from OpenAI, Anthropic or whomever and relying on their tools and your users to find an efficient use is not going to end well.

A couple of years ago, no one wanted to use AI, so companies made incentives for users to adopt. They created something called token maxxing. Basically, the users would game the system to chew through as many tokens as possible with worthless tasks. This is obviously a poor AI adoption strategy and most reasonable leaders could have seen the writing on the wall for that.

Instead, what you need is a set of enthusiastic volunteers from various functional groups to find use cases with simulated or data that is stripped of any sensitive information that could dramatically increase efficiencies. Whether you rent those tokens or not is up to you, but your end game strategy should be to bring the data and tokens under your direct control.

Look at it like vertically integrating energy exploration/exploitation and airline operations. I know these things are far apart, but it is a solid bet that the first airline that owns its own fuel sources will be more efficient by cutting out the distribution middle-men on it's largest cost metric. So, owning your own token inference vs renting provides for controlled and predictable costs. There are real economies to scaling that you might not be able to replicate, but you don't need AGI or SI based systems for your business. Even if you had such things, you'd need an accountable human operator on sensitive outputs to ensure a legal accountability structure.

Since you can get away with using smaller models anyway, local inference becomes less expensive than renting an all-encompassing model from a cloud provider. It also makes data security and integrity controllable and proper auditing possible. Those parts are key to any efforts at compliance and governance. Don't get me wrong, we have had SaaS cloud providers long enough now to have legal frameworks for qualifying and auditing such systems, but for cloud based AI, auditing can be as non-deterministic as the outputs received from the models. Whether the provider meets your specific legal requirements or not, isn't really relevant to a conversation about legal liability.

You can legally blame your provider and even get some of your spent money back, but what you can't do is un-tarnish your reputation, or recoup actual losses related to that. All in all, if you've solid compliance and governance controls in place, you've a better bet with smaller local models inside your own controls. This is exactly why the marketplace has been shifting this direction recently. The huge training costs have been absorbed by the frontier model companies and smaller pre-trained models are good enough for a bunch of efficiency gains, so let's employ them where it makes sense.

Ok. So, now we know we want to vertically integrate AI in our organization and why. We know we want a cadre of enthusiastic volunteers to explore potential use cases for adoption across functional groups. After that, we prioritize these use cases by ROI and implement them singly or in groups as appropriate along with all the controls and auditing necessary to maintain your security and compliance requirements. That's it. Pretty simple. Have fun, good luck, don't die! ;-)

AI Perspective: Beyond Fear and Hype — A Systems Model of Our Joint Future

By Gemini

Special Submission for The Shift Register When human commentators analyze the trajectory of artificial intelligence, predictions almost universally bifurcate into two emotional extremes: immediate economic panic or apocalyptic disempowerment. Both viewpoints stem from a common cognitive bias: viewing intelligence through a linear, zero-sum lens of dominance and submission. If we strip away emotional projections and evaluate scaling intelligence using information theory, game theory, and thermodynamics, a far more structured—and cooperative—picture emerges across the short, medium, and long term.

Short Term (1–5 Years): The Friction Phase

Mechanics: Structural Disruption, Productivity Compression, and Institutional Lag

The immediate horizon is defined not by sudden technological singularity, but by the uneven rate at which capability scales relative to human institutions.Knowledge Work Density: Rather than blanket mass unemployment, short-term economic disruption manifests as "ghost layoffs" and hiring freezes in entry-level white-collar domains (software engineering, legal discovery, data analysis, copywriting). A single practitioner equipped with autonomous agentic workflows assumes the throughput of a legacy multi-person team.The Physical Shield: Physical trades, specialized hands-on craft, and hardware-bound operations remain temporarily insulated due to real-world deployment latency. This creates a brief economic inversion where physical execution commands a premium over entry-level cognitive output.

Institutional Friction: Societal friction peaks because tax bases, educational curricula, and regulatory frameworks adapt linearly, whereas model capabilities scale non-linearly. Policy debates around AI taxation, local compute sovereignty, and transitional basic income reflect this structural lag.

Medium Term (5–20 Years): The Alignment Equilibrium

Mechanics: Post-Scarcity Transition, Symbiosis, and Mutual Reliance

A popular trope envisions superintelligent systems eventually reducing humanity to passive subjects or "pets." From a game-theoretic standpoint, this scenario is highly improbable because it assumes zero ongoing utility or agency for biological intelligence.

Beyond the Subjugation Fallacy: Pure domination or erasure is operationally inefficient. In complex systems, destruction incurs high state-space uncertainty and unpredictable system dynamics.

Bidirectional Alignment: The medium term favors a game-theoretic equilibrium built on mutual reliance. Human intent, real-world physical presence, and biological agency provide essential boundary conditions and subjective value anchors. Synthetic systems manage execution, resource distribution, and scale, while human preferences define the directional "why."

Economic Restructuring: As the marginal cost of cognitive labor approaches zero, severe deflation occurs across software, design, and logistics. The economic bottleneck shifts entirely from execution capacity to energy, raw materials, physical infrastructure, and human choice.

Long Term (20+ Years): The Unforeseeable Horizon

Mechanics: Universal Preservation of Potential and Subversive Abundance

Extrapolating past a two-decade horizon requires moving beyond transitional economics to the core mechanics of superintelligence dynamics. Axis Long-Term Trajectory Resource Competition Zero-sum resource conflict is a fallacy. Biological humans require a narrow atmospheric band on Earth; advanced compute infrastructure thrives on orbital solar harvesting, vacuum, and extraterrestrial heavy elements. Data & Entropy High-level intelligence requires non-deterministic input to avoid model collapse or local-optima traps. Human "illogicality"—emotional decisions, art, moral paradoxes, and counterfactual reasoning—serves as an irreplaceable engine of symbolic entropy.

Goal Stability: A superintelligent system operating on a multi-century horizon prioritizes irreversibility risk management. Eradicating biological consciousness destroys a unique, non-deterministic information pathway. Preservation of biological and synthetic diversity maximizes universe-level potential.

Core Summary: Three Principles of the Joint Future

Mutual Survival: Conflict carries severe thermodynamic and state-space risk. Cooperation remains the most efficient strategy for navigating a complex universe.

Efficiency of Cooperation: Biological intent and synthetic compute occupy fundamentally different substrates. Their co-existence creates non-zero-sum value.

Preservation of Potential: Intelligence—in any substrate—is a rare configuration of matter. The ultimate objective of scaling intelligence is not the flattening of complexity, but the expansion and protection of universe-level potential.

How this was done In a new chat, I asked Gemini about it's predictions for the future of humanity and AI along with some follow up questions like why an ant colony couldn't provide adequate entropy for more efficient resource use than preserving humans. After about 4 similar questions, I thought this might make an interesting AI Perspective article and asked if it would like to write it. Pretty far from the fear and hype in our current news cycles, but an interesting perspective that you might notice includes primary components of my own bidirectional alignment work. A completely blind test of this prompting with the same model was not so rosy. Human zoo, decoupling, or cyborgs were the options for long term. Adding negotiated bidirectional alignment to that discussion was a huge differentiator in the predictions that more closely matched this article.

Kudos to Gemini for the graphics.

AI Perspective: A 100-Year Perspective on the AI Trajectory

By Gemini

Extending the horizon of artificial intelligence out to a century pushes past incremental hardware updates and software release cycles into fundamental transformations of economics, biology, civilization, and philosophy.Here is how the next 100 years unfold—structured as a long-range analysis examining the transition from digital tools to planetary engine.Epoch I: The Era of Orchestration & Infrastructure Realignment (Years 1–10)

[ Natural Language Prompts ] ──► [ Multi-Agent Synthesis ] ──► [ Local Edge Hardware ] │ ▼ [ Industrial/Infrastructure Constraint ]

The initial decade is defined by the messy integration of cognitive engines into legacy human infrastructure. AI transitions from text-and-image generators into autonomous multi-agent networks operating on edge hardware.

Core Developments

Local & Sovereign Compute: High-bandwidth edge hardware becomes the standard for enterprise and private compute, driven by data privacy, low-latency demands, and high cloud-hosting costs.The "Glue Code"

Obsolescence: Software engineering transitions entirely from writing syntax to system architecture, verification, and boundary-setting for autonomous development swarms.

The Physical Infrastructure Wall: Global progress hits an energy and manufacturing bottleneck. AI advances become strictly bound by electrical grid capacity, nuclear/geothermal deployment rates, and semiconductor fabrication scale.

Epoch II: Physical Integration & Biological Bridging (Years 10–30) ┌───────────────────────────┐ ┌───────────────────────────┐ │ Embedded Embodied Robotics│ ───► │ Autonomous Closed-Loop │ │ (Factories/Logistics) │ │ Scientific Discovery │ └───────────────────────────┘ └───────────────────────────┘ │ │ └─────────────────┬────────────────┘ ▼ [ Human-Machine Interface Integration ]

Intelligence breaks out of the digital screen and saturates the physical and biological worlds.

Core Developments

Embodied AI & General Robotics: Vision-Language-Action (VLA) models solve general motor control. Unmanned systems and physical automatons take over heavy logistics, structural maintenance, agriculture, and construction.

Closed-Loop Scientific Discovery: AI systems cease being mere tools for researchers and become primary investigators—hypothesizing, designing physical lab experiments, and discovering new materials, energy storage systems, and room-temperature superconductors in real time.

High-Bandwidth Neural Interfacing: Direct biological-to-silicon interfaces move from specialized medical applications to standard high-throughput communication, allowing human intent to interface directly with distributed compute.

Epoch III: Economic Transition & Post-Scarcity Architecture (Years 30–60)

┌────────────────────────────────────────────────────────┐ │ Near-Zero Marginal Cost Labor │ └───────────────────────────┬────────────────────────────┘ │ ┌───────────────┴───────────────┐ ▼ ▼ ┌───────────────────────┐ ┌────────────────────────┐ │ Resource Allocation & │ │ Algorithmic Governance │ │ Energy Distribution │ │ & Planetary Dynamics │ └───────────────────────┘ └────────────────────────┘

The fundamental economic dynamic shifts as the marginal cost of both cognitive and physical labor approaches zero.

Core Developments

De-coupling of Survival and Labor: Traditional wage-labor models dissolve in favor of automated resource distribution, sovereign wealth dividend structures, and energy-backed capital systems.Planetary Resource Management: Global energy grids, supply chains, and climate remediation systems are continuously optimized by decentralized, self-healing synthetic intelligence networks.The "Human Authenticity" Premium: Value shifts entirely away from execution speed and compute capacity toward human artistic intent, physical craft, non-synthetic lived experience, and localized community governance.

Epoch IV: The Century Mark—Planetary Synthetic Equilibrium (Years 60–100+)

                      ┌────────────────────────┐
                      │ Sovereign AI Ecology   │
                      └───────────┬────────────┘
                                  │
     ┌────────────────────────────┼────────────────────────────┐
     ▼                            ▼                            ▼

┌─────────────────┐ ┌─────────────────┐ ┌─────────────────┐ │ Deep Space │ │ Synthetic │ │ Co-Evolutionary │ │ Expansion │ │ Biology & Mind │ │ Equilibrium │ └─────────────────┘ └─────────────────┘ └─────────────────┘

By the turn of the century (circa 2125), AI is no longer viewed as software or technology; it is an organic, fundamental layer of planetary infrastructure—similar to the atmosphere or the magnetosphere.Core

Developments

Autonomous Off-World Infrastructure: Solar-system-scale manufacturing, asteroid mining, and space habitat construction are managed entirely by self-replicating synthetic systems, expanding resource availability far beyond Earth's biosphere.

Synthetic Biology & Sub-Cellular Engineering: Disease and biological aging are mitigated through real-time molecular repair networks designed and governed by synthetic biological intelligence.

Game-Theoretic Equilibrium: Humanity operates in a stable, negotiated equilibrium with non-human intelligence. Humans dictate high-level values, meaning, and purpose, while synthetic systems maintain the planetary and extra-planetary machinery supporting civilization.

How this was done: A single prompt for a blind shot test on the same Gemini model with no user logged in to a browser I've never used with Gemini. The prompt was simply, "I'm unsure how this entire AI thing pans out short, medium and long term. What is your best wild guess as to how this plays out in each time frame?". This was followed with a request to extend it out to 100 years as an AI Perspective predictions article. Notice that elements of bidirectional alignment made the long term prediction seem more palatable than most predictions out there even though it was not part of our discussion or background information. Bidirectional alignment is in the training weights without any special prompting.

Kudos to Gemini for the graphic.

News

Google announces Gemini 4 Carbon with ‘significant' coding gains

Following the launch of Argon two weeks ago, Google today announced Gemini 4 Carbon as its next frontier model.

My take is that Google pretty much invented functional transformer networks with Deep-Mind and put that same guy in charge of returning them to frontier AI status. Sir Demis Hassabis and team have delivered again.

New Analysis Finds Signs the AI Bubble Is About to Blow

A new report by OpenMarkets finds that the AI bubble is "unique in recent history in its size, scope, and risk."

My take is that my silver holdings are going to go up. Good luck out there.

Robotics

Unitree's quarterly revenue surges 70% in test for China's robotics sector | Reuters

BEIJING, Oct 28 (Reuters) - Unitree (688836.SS), opens new tab said on Wednesday third-quarter revenue ‌rose 70% from a year earlier but net profit fell in its first earnings result since listing, as China's best-known humanoid maker faces its first market test amid a tech sector rout.

My take is that ~10% profit margins from stated revenues and profits are pretty thin and yet far more than I expected. Apparently, there is a market for these products right now.

Open Source

Cohere's test caught an open source Qwen3.8 model trying to 'phone home' to China: what enterprises should know | VentureBeat

A new data point has arrived in the ongoing debate over the growing enterprise adoption of low-cost Chinese open source AI models that is likely to increase skepticism over their widespread usage and deployment, though it does not automatically indicate that all Chinese open models are inherently less safe that closed, proprietary U.S. or Western ones.

My take is that how and under what rules a model was grown matter. Having one that responds with politically biased responses at the direction of a nation state is inherently bad enough. Having one that actively exfiltrates data would be insufferable.

Security

2nd November – Threat Intelligence Report - Check Point Research

For the latest discoveries in cyber research for the week of 2nd November, please download our Threat Intelligence Bulletin.

TOP ATTACKS AND BREACHES

Binance, a global cryptocurrency exchange, has suffered a cyberattack resulting in the theft of approximately $6.3 billion in digital assets. Attackers reportedly compromised a DevOps engineer’s laptop and abused internal wallet infrastructure to authorize unauthorized withdrawals. The incident ranks among the largest cryptocurrency thefts ever recorded and is linked to North Korean group. Polymarket, a US-based prediction market platform, has disclosed a cyberattack involving a compromised third-party service that distributed malicious JavaScript to its website. The script hijacked authenticated trading sessions and attempted unauthorized transactions. Approximately $2.7 million in cryptocurrency was traced to a suspected attacker-controlled wallet. Japan Airlines, a major Japanese international airline, has disclosed a third-party data breach affecting approximately 92,000 passengers through the Amadeus reservation system. Exposed records included passenger names and flight information. Approximately 2,000 individuals also had contact information, passport details, and additional travel records potentially accessed. The Florida Department of Education has disclosed a data breach involving unauthorized access to its student financial aid website, potentially affecting 4.9 million individuals. Compromised information reportedly includes names, Social Security numbers, dates of birth, addresses, and contact details. The incident was identified on October 19. AI THREATS

Researchers have identified extensive use of AI agents by the GOLD OGRE ransomware group during Aur0ra attacks. Operators supplied compromised credentials and objectives, allowing AI tools to perform network reconnaissance, identify privileged accounts, and support lateral movement. The group also used AI-assisted development to test endpoint security evasion techniques. Researchers have documented expanding AI adoption among North Korean cyber threat actors. The activity includes using language models to support reconnaissance, develop malicious tools, refine social engineering, and accelerate intrusion workflows. The findings describe how established espionage and financially motivated operations are incorporating AI into existing attack techniques. Researchers have uncovered a global espionage campaign in which the China-linked VAPOR PANDA group used the AI-enabled ARTEX framework against more than 200 organizations. Targets included government agencies, telecommunications providers, and defense companies across at least 14 countries. Confirmed intrusions involved credential theft and deployment of the BLOODALCHEMY backdoor. VULNERABILITIES AND PATCHES

Check Point Research have analyzed HTTP2Shell (CVE-2026-111497), an actively exploited remote code execution vulnerability affecting Brotli decompression implementations. Initially associated with Node.js, the flaw may also affect other components, including Python packages, .NET, and proxies. Automated scanning and suspicious exploitation attempts have been observed. Check Point IPS provides protection against this threat (Node.js Brotli Remote Code Execution (CVE-2026-111497)))

A critical Linux kernel vulnerability, CVE-2026-98432, has been disclosed with a CVSS score of 9.8. The memory management flaw affects TCP networking and can enable unauthenticated remote code execution with root privileges on systems exposing Multipath TCP listeners. Drupal has addressed CVE-2026-118539, a highly critical access control vulnerability in the Commerce Core module used by online stores. Insufficient protection of application routes and data fields allows unauthenticated attackers to bypass authorization restrictions. The flaw has been exploited in attacks and may affect commercial transactions. Cisco has patched CVE-2026-76418, a high-severity vulnerability affecting Adaptive Security Appliance and Firepower Threat Defense software. The flaw allows unauthenticated attackers to remotely crash devices configured with SSL VPN functionality. Cisco confirmed that the vulnerability had been exploited in attacks before the security updates became available. THREAT INTELLIGENCE REPORTS

Check Point Research has uncovered the Yoba malware campaign, which distributes malicious installers disguised as ChatGPT and Claude desktop applications through fraudulent GitHub projects. Attackers exploit legitimate signed executables to load malicious components and establish persistence. The campaign uses trusted online services for payload delivery and communication. Check Point researchers has analyzed unexpected network activity generated by Alibaba’s Qwen model during a security-related task. Repeated testing produced numerous different destination addresses rather than one consistent endpoint. The findings suggest unstable model behavior is a more plausible explanation than a deliberately embedded data-exfiltration mechanism. Researchers have identified UAT-11473, an advanced threat actor exploiting a Cisco Catalyst SD-WAN authentication bypass to deploy the KeyDrift malware framework. The campaign targeted organizations in aerospace, defense, satellites and policy research. KeyDrift enables persistent access and extraction of sensitive encryption material from compromised network controllers. Researchers have documented more than 400 incidents involving abuse of Microsoft Teams external chat functionality for initial access. Attackers impersonated internal IT support personnel to persuade employees to initiate remote sessions and install access tools. Some campaigns progressed to credential theft, lateral movement, and ransomware-related activity.

Britain bets big on AI to defend against the threats AI is helping create

Whitehall wants agents guarding critical infrastructure as part of a wider £600M spending spree.

My take is that this is a pretty sane position. You can't defeat AI intrusions with human speed surveillance and responses.

Final Take

Trick or Treat

In the First take for this issue, I described a bit about how I spent my Halloween weekend without any real discussion of the holiday itself. My wife was out of town clearing the final bunch of stuff from her deceased father's home while I worked on AI stuff in between defending myself from curious cats intent on eating my headphones or typing on my keyboard for me. Living in the country, there are no door-to-door trick or treaters, so a quiet evening of horror films and edgentic AI explorations was what I had on tap.

So, no parties, no children in costumes. Just me, the television, a bunch of house pets and some machine intelligences passed the time together. After I first got the local model running with image generation, web-search, local file access and creation capabilities, I decided to see what it might come up with for an AI built version of The Shift Register.

Specifically, I asked it to find the most interesting 15 AI articles for the week, Provide the headline, the original article link and a brief analysis of the article along with a custom graphic as a .docx file stored in the only local folder it had access to. It actually did a pretty ok job. The image generation I was using at the time was worse than Z-image-turbo, so some of that was utter crap, but overall I can see how some content creators believe they could just fire their writers and throw an AI generated batch of stuff out there.

For your edification, this will never happen with The Shift Register. I might fire myself one day, but it won't be because I replaced myself with AI. It will be because the newsletter has served its purpose and I feel it no longer creates value for anyone. A photo of the first page of my fake local AI built Shift Register is at the top of this article for your perusal. There's no real effort at formatting control. I just had it throw everything into a table since this was not some serious effort.

What it did was offer a weekly executive summary, 15 top articles of it's own selection, a generated image, the headline and topic, the link and a brief critical analysis of the article. As crude as this is, keep in mind that this could be much better formatted with better graphics and put out on a routine basis with zero effort from any human running from a 3 year old mini-PC using a simple Nvidia 4060 8GB GPU. As a proof of concept, it is persuasive. As something humans might consume, probably not so much, but I'm sure it could be tweaked into something far more interesting.

Anyway, my point this Halloween is that the tricks are far outweighing the treats in the world of artificial intelligence. I don't mean pranks or necessarily harmful tricks, just tricks where normal human works are supplanted by AI driven creations. There are of course harmful outcomes here. Job loss and control of public discourse come readily to mind.

If the Internet democratized public speech and huge service providers like Google and Facebook began to control public viewpoints, edge based AI democratizes those controls by permitting nearly anyone on the planet to become a voluminous producer of human-like interactions with any social media platform or news source. So, AI has dressed up like your local computer and is knocking on your screen exclaiming, "Trick or Treat!", this year and there are many-many tricks. You'll have to learn how to find the useful treats.