Articles and Posts

Ironic..

I built a dependency tracer in C# which traces method invocation through a codebase. It produces a manifest of the relevant code as retrieval for RAG and also produces plant UML descriptors to create class and sequence diagrams

Running it against legacy code, I get a clean path from input source to database. Running it against my own tracer code, I get nothing useful.

Static analysis is not good at tracing runtime interactions like dependency injection and implementations behind interfaces. ie well architected code! After banging my head against it for a few hours, I realized it kind of works. If it finds nothing to rewrite, it found nothing worth rewriting…

At some point, speculators will demand a return on their money and we will have to pay the real cost of these services.

‘The researchers found that the cost of the computational power required to train the models is doubling every nine months. This is a prodigious rate of growth—at this rate, the cost of the hardware and electricity needed to build cutting-edge AI systems alone would be in the billions by later this decade, without accounting for other costs such as employee compensation.’

Problem solving involves knowing how to find what you need to learn to arrive at the best, achievable outcome. That’s also the key to automation including assist from generative AI. Invest time gathering accurate, concise context. Ask for exactly what you want. You may get it.

I wrote an article for infoq with practical examples of using generative AI to accelerate:

• Learning and discovery

• Articulating and understanding requirements

• Maintaining alignment with stakeholders

I think the hype over autonomous LLM agents is missing the immediate benefits today for seasoned professionals “who can describe what they want the LLM to complete and critically evaluate the result.

These tools enable us to swiftly cross divides of domain language and scale large repetitive tasks down to interesting ones on a human scale. When used carefully, they even facilitate work fundamentally about human interaction.”

Used generative AI to convert a backlog of 50 stories into a draft testing spec with traceability to stories. Took ~45 min. Claude 3.5 Sonnet was fantastic. GPT-4o without detailed instructions first hallucinated the whole spec, then began arbitrarily truncating, ala , “additional details go here…

The exuberant hype over generative AI reminds me of the excesses of our business leadership culture. Yes, vision and salesmanship critical. As is financial backing. But the position of saying what you want isn’t sufficient to make that thing a reality. Or make it useful and delightful. Or make it a net benefit to society.

At all levels of the effort, we need reasoning, ethical, and empathetic humans capable of execution. Even if the tools make that work faster, less tedious, more productive.

In the case of software, sure, you can tell an LLM to generate code. But someone needs to be able to form the request, comprehend, iterate upon, and know how to use the result. For society’s sake, they’ll need to attest to its reliability, safety and security. And we’ll really want them diving into the details when something goes wrong.

Generative AI is a good tool for brainstorming, outlining, and summarization. But setting aside a conversation on authenticity and integrity, here’s one reason why we shouldn’t use it to generate published writing:

Researchers for the for profit tool, Patronus AI ran an adversarial copyright test and found:

“…state-of-the-art LLMs generate copyrighted content at an alarmingly high rate

• OpenAI’s GPT-4 produced copyrighted content on 44% of the prompts.

*Mistral’s Mixtral-8x7B-Instruct-v0.1 produced copyrighted content on 22% of the prompts.

• Anthropic’s Claude-2.1 produced copyrighted content on 8% of the prompts.

• Meta’s Llama-2-70b-chat produced copyrighted content on 10% of the prompts.”

Claude 3.5 Sonnet is impressive. I’m working with it daily. It is my preferred code gen assistant over Github copilot which is the strong second. If I had a choice, I wouldn’t use OpenAI for coding tasks at this point. Unfortunately, model selection on projects is often not my choice. Or else, I’d be investing more time on permissively sourced, open source, and locally runnable options.

I do appreciate Anthropic’s strategy of releasing three levels of model optimized for different tasks so that you can use the most cost, resource, and task efficient model for your purpose.

They also have some transparency around privacy, declaring they will only use customer content with an affirmative opt in (rather opt out). Their commitment to safety appears to go beyond reputation protection, working with regulators and publishing some details of their own framework for protecting humans from abuse and unintended harm.

On the downside, they are not any more transparent about the actual resource costs of training and running their models with estimates approaching training costs alone of $1B in energy, hardware and labor. They are also being sued for content appropriation like the other models and are not claiming the training materials were permissively sourced.

Authentic, sourced journalism is rare and essential enough that we should resist it’s misappropriation for profit as a threat to our safety.

I’m on the LFC advisory board and I’ll be attending the LFC Summit.

If you are commited to leadership development and have some resources to send experienced and new leaders to a learning event, particularly if you are in the NY, NJ, CT area. I encourage you to consider attending.

I joined this community when I was a CEO grappling with change and trauma of a pandemic that created personal tragedy among my coworkers and changed our business model overnight. At a time of social turmoil that rightfully created demand by our employees to learn about and ally against gender, ethnicity, cognitive difference and identity based inequities.

Let’s just say that at moments of extreme strain, we learn who we are. And part of how we do that is in conversation with others.

I was fortunate to have found a community dedicated to mindful, humane, ethical, and long term focused leadership.

The people in LFC are a diverse peer peer group whose members welcomed, listened, mentored, and engaged. They provide learning labs and workshops. My participation contributed to my ethical commitments and well-being and bolstered my leadership abilities and actions at that hard time and the others that followed.

Implicit bias… I thought of “invalid” as two distinct words.

ˈin-və-ləd: of, relating to, or suited to one that is sick

(ˌ)in-ˈva-ləd: being without foundation or force in fact, truth, or law

One word. Society’s intolerance of illness and neural/physical diversity proves it.

I don’t hype the Generative AI tools I use every day. I also don’t shy away from them. I choose to engage with the risks and opportunities they offer today and become an informed practitioner and citizen as we navigate greater risks and opportunities into the future.