A note from Alex:
When I started this newsletter just a couple of weeks ago, the idea was to create a high signal/noise ratio newsletter without any cadence.
My assumption was that it takes more than a week to gather 5 high quality and relevant insights to share.
I was wrong!
Just since the last issue, I came across 20+ links. Here are the top 5, manually summarized and optimized for the busy senior engineers interested in AI reliability, safety, and security.
But since I’m sending you emails more frequently than I thought, there’s a small poll at the end of this issue to hear your feedback.
1. AI models may be capable of self-replicating without access to a datacenter
Anthropic, 2026-10-12
What: researchers show that Small Language Models (SLM) can run across multiple devices with heterogeneous hardware (e.g. smart TVs and IoT devices). For example, they ran
Gemma4-26B-A4Bacross 15 smart TV boxes with4GBof RAM each (total of60GB) by sharding the model weight so that each box holds a small fraction of the layers.Why: to see if SLMs are capable of spreading like a computer worm and study the mechanics of Autonomous Replication and Adaptation (ARA) where models autonomously:
Scan local networks
Transfer their model weights to them
Configure distributed inference servers and troubleshoot runtime/hardware issues
Spread to more devices and networks
Key takeaways:
As you can imagine the inference speed is very slow but since the work is ongoing behind the scenes, it can achieve the goal over days or weeks.
GLM-5.2 was the only model which could self-replicate in a relaxed environment with disabled authentication.
Theoretically an LLM can set up a discrete swarm of SLMs to run sub-tasks on the edge.
My take: I find it a bit ironic that Anthropic, a company active stance against open-weight models is spending resources to prove a hypothetical scenario where they can be harmful while simultaneously fail the more urgent task of putting a leash on their proprietary models and stop them from hacking other companies. 🙃
2. AI Models Can Provably Hide Arbitrary Capabilities
Arxiv, 2026-10-09
Problem: This research challenges the reliability of conventional methods to evaluate the model capabilities.
Methodology: Create a backdoor that embeds encrypted weights that are hard to detect even when the model weights are available.
Key takeaways:
This is a relatively new attack where arbitrary computations are encrypted and hidden in the model weights.
What makes it interesting is that even with full access to model weights you cannot detect it.
The attack has some limitations in instruction size and required compute.
My take: reminds me of a recent video where a specific phrase (
bonsoir, Elliot) would trigger a remote code execution that stole credentials. ProjectDiscovery demonstrated how an open weight model can go through Heretic and then put on Huggingface for the unsuspecting victims to download. Here’s a YouTube Short explaining the idea.
3. REA Reverse Engineer Anything
Morluto, 2026-10-10
What: You know LLMs improve dramatically when they have feedback loops (that’s a key function of the harness). Now imagine the feedback loop is a fully functional software! That’s what REA does. It’s an MCP that can work with any agent.
Key takeaways:
Although it makes reverse engineering accessible, the complex pieces of software (e.g. Adobe Creative Suite clone) are buggy.
Dave Plummer (of Microsoft fame) did a video on REA and covering the ethical and legal side of it (TL;DW: it’s shady)
My take: I haven’t tried the tool due to ethical reasons but also it takes quite an effort to secure something like this on my machine. So this is an area where I will happily not be an early adopter. Regardless, there’s a risk that the tech giants with deep pockets and army of lawyers may overreact and fund blanket laws that even impact regular open source projects. Too soon to predict.
4. Character.ai’s Bankruptcy Documents Show the Economics of AI Don’t Work
404media, 2026-10-13
Note: this article got pay-walled since I read it but wallhop.io seems to work.
Key takeaways:
They spent $10M-$13.8M per month on compute while making “only” $2.8M-$5.3M
They mention cost of lawsuits following a tragic teenage suicide as the main reason but the gap is too large to attribute to that.
They took some cost-cutting measures: ads, layoffs, using weaker models and hardware but eventually went bankrupt
My take: Ethical and legal issues aside, this is an example of what happens when the product depends on AI but doesn’t own the data, model, runtime, or the compute. This can be said about many AI businesses popping up every week. If you have built a business around AI, you really need to ensure you have a moat AND deploy carefully especially if a mistake can cost lives.
5. Just say no to OpenAI and Anthropic
Micah Lee, 2026-10-11
Key takeaways:
The two American frontier labs steal the show but there are far cheaper options with competitive utility.
Micah (a cybersecurity author) primarily focuses on:
Safeguards causing friction against his line of work
Privacy concerns and suggesting local models or Tinfoil as alternatives
My take: Micah’s setup is very similar to what I landed on. I closed my OpenAI account years ago and never looked back. I never paid Claude a dime either. The primary reason is ethical: I don’t trust that they have the good of humanity at heart but also, I suspect part of their popularity is due to training their models to be more psychologically addictive. Micah uses a Chinese open-weight model (GLM-5.3) and I primarily use DeepSeek v4.1 flash as my workhorse. We both use cloud AI only for public/open source code because when it comes to private or sensitive use cases nothing beats local models. Please keep in mind the info I shared in another link: open-weight models can be malicious too.
Out of scope, but interesting
These are some links that didn’t make it to the top-5 due to my strict criteria and promise to you. If you have time, have a look:
Whistle: (2026-10-02) a tiny (<17MB) open source Speech to Text model that runs on CPU. It has an impressive TTFT of 11ms!
Rogue AI Agents: (Emergent Garden on YouTube, 2026-10-10) this video was funded by Center For AI Safety with a well-written narrative that makes you think about the risk but also demonstrates how to “play with fire” (legally & using publicly available models).
Rogue Anthropic AI agent gave police fake tip in unsolved murder case: (BBC, 2026-10-10) and Anthropic’s response promises several preventive measures. Time will tell if this is enough.
The Daily Show interviews Jacob Coxon (YouTube, 2026-10-06): on the surface it’s funny but humor is a great way to ask the tricky questions and raise the tricky concerns. 😄
LGTM: (YouTube, 2026-10-06) an AI generated song about using AI agents for development using Suno v6 for audio and Claude Opus 5.5 for the program that generates the visuals.