First Newsletter.

Automation, reasoning, AI tools and the race to win.

Author

Mike Hall
February 05, 2025

A bot for that. ab0t.com - Secure AI App Store

Automation

The rise in interest of automation tools, is a curious side effect of AI hype.

Something previously seen as 'technical' or 'boring', is getting a rebrand away from the previous 'difficult' label to 'possible'.
New no code tools like Make, n8n, Relay - and other tools like Zapier. These tools offer a visual interface to set up simple workflows with existing services you already use. We have seen more and more non technical, to semi technical people start to experiment with building with these tools, and think this trend is only going to continue.

Our suggestion, pick one, try it out for a month - See what you can automate.

A unique moment in social media exists where professionals are socially rewarding technical thinking.

ab0t: Monthly AI Tool Suggestions

Tools are popping up all the time, our suggestion pick two try them out for a week.

Share your thoughts about these tools on your social media of choice.

Here is some of the top AI tools we have seen trending for professional applications:

  • Sales: Clay

  • Presentations: Gamma

  • App Builders & Coding: Bubble, Bolt, Cursor, v0

  • Video Generation and Editing: Synthesia, Runway, Filmora, OpusClip

  • Knowledge Management: Notion AI Q&A, Guru

  • Transcription and Meeting Assistants: tl;dv, Nyota

The Data War

A trade war is a tit-for-tat exchange of tariffs, sanctions, and policies meant to weaken another country's economy, boost local industries, or destabilise rival nations - a mercantile focus on trade balance. Similarly, a data war is emerging: DeepSeek's release has accelerated a years-long trend where public datasets (used to train AI models) are being restricted or removed. Resources like Wikipedia, LibGen, Reddit, Hacker News, Common Crawl, arXiv, and Archive.org - once freely accessible to researchers - are now harder to access. Some are locked behind paywalls (e.g., Reddit monetising its data), others face legal scrubbing.

This mirrors the decline of old machine learning datasets: niche training data, broken Google Drive links, dead websites, and paywalls have erased much of Kaggle's original repository (once a hub for models and datasets). Hugging Face has become the new Kaggle - but our advice is to archive everything you can now. In the coming data war, critical training data for future AI models will vanish or be sanitised from existing sources, this will happen slowly, then all at once.

Archive huggingface.

Reasoning: I think therefore I am?

The latest saga in the path to AGI is simply appending, ‘wait,’ every few tokens, during the thinking phase to force the model to reconsider its previous possibilities, while funny this simple approach works, it takes a step deeper away from typical LLM "blurting" answers, to still blurt, but then to take a moment and conciser if that blurt was valuable. Maybe similar to what happens in your mind, before you speak aloud.

Reasoning also has a UX benefit for consumers, users can follow the train of thought, which seems to humanise the experience, they can notice when things go in the wrong direction, Identify assumptions they need to clarify.

Most of good prompting is being unambiguous as possible, removing all possibility for assumptions.

YC: Request for startup

Ycombinator put out its call for startups to join their accelerator, Spring 2025. Applications end Feb 11th, here are a few from the list.

  1. A Secure AI App Store: A privacy-focused AI app ecosystem with user-controlled data, shared memory, vetted tools, and seamless payments for smarter, private AI apps.

  2. AI-powered document platforms: That auto-fill, explain terms, and customise templates for frictionless e-signatures.

  3. Browser & Computer Automation: AI agents that automate workflows by interacting with websites/apps as "virtual APIs" for limitless use cases.

  4. AI Personal Staff for Everyone: Personalised services (e.g., legal, financial, health) via AI to replace roles only the wealthy afford.

  5. Devtools for AI Agents: Tools and APIs to build, deploy, and enhance autonomous AI agents for productivity and specialised tasks.

  6. The Future of Software Engineering: Platforms enabling developers to manage AI agent teams for coding, testing, and deploying software at scale.

  7. B2A: Software Where Customers Will All Be Agents: Services designed for AI agents as primary users, with APIs for payments, contracts, and interactions.

  8. Vertical AI Agents: Domain-specific AI agents (e.g., tax, healthcare, compliance) automating entire workflows to replace human roles.

ab0t: Secure AI App Store

This site, ab0t.com is infact the number 1 thing they are looking for (still under development). A major and open problem in AI agents, is firstly finding them, and then access control, let them have access to your system, your apps, your gmail, calendar, google drive.

Most app exploit this and try gain as much access as they possibly can, "Full google drive access", when really they only need access to a specific file in a specific folder for 24 hours.

This fine-grained access control is such a major security concern, its the default way both iOS and Android do things. Apps are restricted by default, and can only do very specific actions the user allows them access to. Where a user can manually toggle a apps access, rather than granting access. This and a host of other reasons, like shared APIs, payments, vetting, human in the loop, discovery, autonomous discovery of services by AI agents, make this a very big problem to solve.

Companies like /dev/agent have raised 56 million seed round on this thesis, “IOS of agents”. Where agents, workflows and APIs built on other services, can be shared by developers and sold to other people, on a secure platform.

AI assistants: Seri on steroids

AI assistants, this we believe is the most obvious mass consumer use case. A 'friend' you can talk to with your voice, and can do things for you, like seri on steroids. Many already use chatgpt in this way, rather than using the website, they have the app on their phone. Meta announced in their recent earnings call, they aim to win this category, and a number of voice enabled AI assistants are popping up. We believe this is an interesting trend to keep an eye out for. Mass market consumer AI will be via this medium.

Voice enables a open playing field

Previously, voice models were either too expensive for this to be a valid feature for the consumer market, or large companies created private models that could be run at a low cost, now developers and startups have a reasonable quality opensource version that is going to open up this space, eventually being able to run these models on the phone.

The kororo model, just reached version 1.0, trained on $1000 budget from a base model. Plenty small to run on even old devices.

Here is a sample from the kororo model for reference.

Interesting papers & articles.

LLM autonomously creating backups, ie self-replicating

Frontier AI systems have surpassed the self-replicating red line - "Replication is the process of making an exact copy or duplicate of something. It can refer to the process of copying DNA in cells, or the process of repeating an experiment"

Replicate the Deekseek reasoning training

Papers

Articles

Linkedin