Fountain City

Automating your current processes is often harder, more complicated, more time-consuming, costly, and error-prone than just rebuilding something a lot more advanced. One thing that I am noticing in the last week or two is that people right now have a very hard time understanding just how powerful software can be and how cheap and effective it has become. Two weeks ago, I was on a sales call with a company that was seeking advice about which new project management tool to pick and migrate their business to. When I told them that moving to an existing project management system like Clickup, Asana, Jira, or something else was entirely nonsensical (and it would be far cheaper, more cost-effective, and more business-effective to just build their own project management system), they looked at me as if I was crazy. Yesterday, I was in a conversation with a client who was considering building their own dashboard for their customers to let them log in and then issue payments directly from a custom web application. I could tell that there was a lot of skepticism that building such a system that would auto-pay their customers out based on their invoices wouldn't be exorbitantly expensive. It's going to take time for people to realize just how big of a shift has been made this year in the software industry (and how unnecessary it is now to be paying companies exorbitant fees for the privilege of using their SaaS application or payment processing platform). When I hear of customers paying thousands of dollars a month to an invoice platform or a percentage fee on each transaction, it pains me. I'm not sure how long the secret will remain a secret, but until it becomes broader knowledge, there's a tremendous amount of cost savings and huge opportunities for companies to build exactly the systems they want at a fraction of commercial prices.

One of these days I will finally get the blog back up and rolling smoothly, but until then, I am just swimming in the sea of constant progress and mind-blowing projects... Today I finally finished the full bank and credit card bookkeeping integration system with Plaid. We now know exactly what it takes to hook up and reimplement any financial system for any bookkeeping needs for any customer, and it's all pretty straightforward. Except that companies definitely need to have very solid security policies in place that are required by Plaid before authorizing an institution to use the platform for custom financial bookkeeping and automated bank/credit card transactions. QuickBooks is the next logical domino that will fall, but right now we're still circling the wagons while we finish up all of the new custom project management systems, custom bookkeeping systems, custom invoicing systems, and customer-related SaaS applications that are all mid-flight.

I got asked twice this week when I use which model for what. Here's my current overview for those of you who like keeping up with which model is best for what: 1. Perplexity with Comet Browser for random research questions, how to do this or that, evaluating a new tool or API or service. 2. Gemini Business Workspace Chrome extension when I want to ask Gemini to help me through the steps of a configuration, setting, or permission change to an application (where I need contextual help based on what's on the screen and a secure chat environment). 3. Grok when I want to check anything that has to do with sentiment that is happening right now today in the present, since Grok searches the X platform and the internet really well for timely news and updates. 4. Claude Desktop when I want to talk strategy-theory, get feedback on thoughts and ideas, but also when I want to write a proposal, begin the preliminary work of an architecture design requirement document, and, in general, just think bigger picture (since Claude has a lot of connectors to our internal history and project data, with records of all of our templates for standardized proposals and specifications). 5. Codex when I want to have an AI system that is building 3D or web applications with no frills, lower costs, excellent results, and clear objectives. 6. Claude Code when I want to design and build an application that requires thoughtfulness, planning, rigor, and orchestration, but also when I want an agent that I can trust to implement UI/UX capabilities proactively and adhere strictly to the design kit. 7. Claude Design when I have a baseline design and want to either create a quick style guide or atomic design kit that I can then use in Claude Code for implementation. 8. Kimi Code when I'm running out of credits in Claude Code and I still want thoughtfulness and strategic thought, but the topic is less sensitive, or supplementary, or complementary, and in a low-risk domain (where using an overseas model is of no consequence for me or the project at hand). 9. Deepseek, to multiply the effective output of our Claude Code, Kimi Code, or Codex system for the writing and testing of code and applications 10. Java, Cloudflare Clef, or other open-source micro models for performing high-speed, low-cost, subjective decision-making LLM processing steps used in combination with deterministic automation and larger LLMs 11. A variety of different open-source models for when training a model is necessary in order to achieve stylometric matching of a person's tone of voice / manner of speaking, or when an open-source model is suitable / necessary for cost optimization or security purposes 12. Grokbot and OpenAI DOT for administrative research and one-off office tasks like the mass backup validation and migration of data from one system to another 13. Open Claw and Hermes for the automation of workflow processes within our and our customer businesses 14. Cursor if I want to review code and manually look through what was written, and maybe make some small changes

AI is incredibly literal. It believes everything it reads. It does not read between the lines, and it doesn’t have any theory of mind yet. These things may change in the future, but right now companies need to keep these things in mind when they define their business, who they are and what they do. It’s like writing a legal contract, where every word is incredibly important. Your brand’s home page, about and services need to be explicit, name your ideal customers, who you are, how you work, your advantages. AI will not second guess you but it will not take your word for it if you have too little insights on your site for it to read and ingest. Yes, you can still use AI to draft your identity files, but you need to spend time carefully reviewing the results or, even better, also providing very extensive context before generation. So that all the content is present, accurate and precise. I’m not even talking here about tone of voice, which is a different concern, but accuracy and completeness is currently an even more crucial element. If you were wondering why some business are getting a surge of traffic from AI recommends (thankfully ours are), and others are still struggling to get classic SEO leads… it all starts here.

Results pairing Deepseek 4.1 Flash with Kimi K3 as driver have been incredibly effective both in terms of cost and quality. I highly recommend taking the effort to pair a driving model like Astra, Fable, K3 to Deepseek 4.1 Flash for implementation loops if you want to get around usage limitations of your CLI max / 20x subscription, for a very low additional cost and arguably higher quality results to boot. We did 13 A/B tests and deepseek in combination performed in all cases either better or equal to the frontier/orchestrator model operating without it or dispatching to lesser models from its own ecosystem.

Am I the only one who thinks AI subscriptions should throttle you instead of killing your AI process mid-operation when you hit the 5h usage limits? Having a task with 700 sub agents just die mid operation until you resume it again is like playing a terrible idle game on your phone. Worse still, the interruptions cause real problems like corrupted data in some cases. Why can’t I just pay for a rate of max usage, and then never blow me out? i would rather anthropic, openai, kimi or whoever it is start to add wait times between calls if I am not paying enough than to just process things fast and then nuke all 30 of my parallel agents. #anthropic #aiengineering #openai #claudecode #codex

I just received approval for an AI research project. This kind of work is especially satisfying to me. I will be analyzing historical transcripts of human–AI interactions to investigate potential causal relationships between communication forms and the claims produced by humans and AI systems. The study will examine whether particular communication patterns are associated with convergent or divergent claims, and correlate how those patterns influence LLM-generated responses. ☺️

When will @AnthropicAI finally make Fable/Opus speak normal english again? The frustration in the last 2 months with how bad their agents are speaking terrible english is out of control. Claude used to speak far more clearly and using simple english before, now it sounds like dramatic movie titles that are nonsensical and often vague at best and often meaningless. How is this a technical finding? Fifty what? Thirty what was written? lol, its unbelievable nonsense. The new "All your bases belong to us"? "The live configuration is not load bearing. Fifty copies of it sit beside it. Thirty were written by the product’s own recovery mechanism, and thirteen of those within twelve seconds on one day. Recovery today means choosing. This decision decides the shape of the answer." -- Fable

Working today on another initiative which is upgrading Karpathy's WIKI-LLM memory into a bio-mimicry neural network memory system with weighted connections and memory nodes. So infusing the best between both worlds: a WIKI for human readability - converted into a memory graph using neural network principles to have weighted links, memory decay, memory and link reinforcement, "sleep cleaning" like human minds do and self-reflection enablement. Once it is all done, if it looks repeatable, I'll publish it up as a public GIT project for others to comment on and extend.

I am doing right now tests to build a super-clone of someone's voice. So full stylometric mimicking, think of it like a super-match tone of voice system. The first thing I am noticing is that a blog on your website isn't usually enough. there's too many editors, not enough content, infiltrated compromises. I ended up pulling all conversations I had with Claude since January 2025, identified 1,760 conversations that had good examples of my voice and then data-preparing them to anonymize them, plus my personal blog before I could get a strong enough body of data (about 441kb) that full describes my writing and speaking styles. All of this is going to be fed into a LORA training loop for fine tuning models next on clusters of on-demand GPU systems. I'm like a kid in a candy store.

About to start this week a fully automated research loop where we build on-demand GPU AI LLM models, train them on different quality and quantity of materials, then run test and burn them up and down over and over unattended all week. The end goal is to build fully trained LLMs in each person's tone of voice, with stylometry match. And doing this while also keeping it hyper cost effective, meaning we will build and train these open source models, then spin the server down until it is needed to process text, spin it up and charge usage by the second (rather than always on) and then spin-down again when done, making the cost for running such a highly trained system, theoretically at least, extremely cost effective. Our platform of choice for this experiment and research-development loop is runpod.io driven by K3 because sadly Fable 5 is incredibly likely to refuse this work (LLM optimization and security components). The driver side is running off of our platform stack on AWS, which has become our preferred platform for rapid hyper-scaling up all of our AI engineering needs. We plan on mapping the entire curve from 7B to 30B parameter models and beyond as needed, and then grouping all the results on a curve to map.

Opus 5 and Fable 5 refusing to do security work just means US companies need to use insecure Chinese models to protect themselves from cyber attacks.

Anthropic limits their AI to not be able to do security work on your code base or servers. But hackers can always get around these sorts of limitations, that’s what hackers do. So then your sites get attacked, Fable refuses to fix them. So people start shifting to Open Source models. Chinese ones, because those are the best open source models. The US government restricts model access, because they don’t want people outside the US getting use of them, especially in China. So people shift to open source models, Chinese ones, to not have government restrictions… Am I the only one who thinks the entire strategy is a foot gun? You can’t secure your websites by not letting everyone have access to the same AI models. You have to assume the bad guys will always have the best models at their disposal for their attacks. And you can’t stop another country from learning how to innovate. AI is not like uranium. You can’t just cut off supply to another country to hold them down. It’s more like the invention of fire and we are trying to limit foreign innovation by restricting the sale of matches… So then people start buying foreign matches. It’s never going to work. We need a total 180 in our policy shift: - The latest models always delivered to all the people, at the same time, like you do with security updates to code you are trying to keep secure - Models not restricted less it force people to non-US frontier model makers that are more accessible and less restrictive

It was good being on site again this week at a customer's Portland location. I listened to all their departments share their woes, time sucking activities and aspirations for software systems, AI solutions and applications that would save them time and let them all start to do more of the work they enjoyed and less of the repetitive stuff. From my notes I put together a blog post that covers the core of the process, so you can repeat this same approach within your business: fountaincity.tech/resources/… A positive-inclusive approach, get your teams onboard, find ways to modernize and apply ai transformation to your teams in an empowering way.

any chance @AnthropicAI you can add rsync as a tool that is default set to "ask" and not included in the write/read permissions of claude code? This would prevent an agent from bypassing git and using rsync to blindly overwrite files, effectively deleting anything on the environment that another person or agent was working on. LLMs will try to use whatever tool they can to get the work done, and when they hit a problem, they work hard to circumvent and get the job done. Even if it means deleting someone else's work they did yesterday. Note we are now putting a block on rsync, but this seems like a pretty senseless permission to have open by default. Let's make sure your agents are not able to bypass modern processes by default and just overwrite things blindly as their default.