Robert (@robert2sj) header
Robert (@robert2sj) profile picture

Olives & Code

Joined January 2014
Making olive oil and writing code.
1
87
Just migrated 3 apps from Render to Kamal. Postgres, Sidekiq, Redis, DB backups to S3. It makes it simple to swap out the VM too.
1
40
Building my first project in Django and Celery for background jobs. Wish me luck 🤞
1
51
I built an app to create meeting summaries. 1) Record your in-person meeting (audio only or video) 2) Upload the recording 3) The app creates the meeting minutes
1
50
You can try it at elacta.com (Spanish interface only atm).
36
Trying crewai agents and they're pretty impressive.
1
46
But they create so many prompts that they quickly incur a lot of costs with OpenAI.
1
28
Now operating them with a local LLM (Llama3) it runs for free. Sweet!
39
One of the biggest questions in AI right now, especially in talking with investors, is where it will have the most amount of impact across the enterprise software landscape. It's critical to first start with what AI is good at and why it is so important. AI --at least in its forms for the foreseeable future-- has the ability to provide a reasoning engine for working on information (text, audio, video, or images, etc.) in service of bringing automation to small or large bodies of non-deterministic work. Critically, however, it's useful to separate which areas of software will have the most amount of potential upside from AI, and which areas will AI be mostly just a nice-to-have improvement. When AI is used to automate a small area of work, like we saw in a mad rush right after the launch of ChatGPT (let's say helping a user navigate software via a prompt interface vs. a GUI), we can imagine that this level of automation has a fairly incremental amount of value. Customers likely won't pay extra for this value proposition, and we likely wouldn't see much lift in customer demand for that particular area of service. Conversely, if we could instead automate an entire type of work in a piece of software that otherwise might take hours of high-priced labor to accomplish, we can see this as being fairly valuable. Even better, is if this is the type of work that happens frequently and has no particular upper-limit to the amount of time that could be spent on the activity. As such, we can imagine at least 3 major axes (and there are definitely more) that will determine the vast majority of value that will be generated with AI: 1. What is the level of automation being applied to the work? The smallest area of work would essentially be an autocomplete in a text field or a chatbot, and the largest area of work would be an AI Agent that automates an entire process or does a substantial amount of research and work to return information back to the user. Both ends of this spectrum can certainly be valuable, but largely depending on how the AI is leveraged and what types of ROI you can generate with the automation itself. 2. What is the "cost" of the work that's being automated? Every task in the economy has a different level of cost associated with it, either because the task itself is very time-consuming or because the level of specialization to complete the task is very high, and thus labor tends to be more expensive. Whether it's writing software code, generating text for a blog post, reviewing a contract, or providing synthesis of equity research, each of these tasks are valued differently. Fairly intuitively, the more that AI is applied to otherwise expensive activities and tasks, the more valuable it is. 3. What is the volume or frequency of the work that's being automated? If it's something that is done somewhat infrequently, like changing the settings of an application via a chatbot, then obviously AI is going to offer a limited amount of value. If, instead, the AI is being applied to a process that is executed hundreds or thousands of times a day in an organization, like QA testing software or routing invoices, then the value is much more meaningful. The pinnacle, of course, would be to go after software categories where the activity is high value, high volume, and offers the opportunity for a substantial amount of automation. Conversely, the worst spot to be would be relatively low cost activity, that is infrequent, and just experiencing the AI through a basic chatbot. But there can of course be different optimizations available: for instance, bringing heavy automation to a high volume low cost activity could be just as valuable as bringing a medium amount of automation to a medium volume, high cost activity. What is clear is that while not all AI opportunities will be the same, and we're only in the earliest stages figuring out which ones will bring the greatest amount of upside, the potential is insane.
45
92
35
687
199,263
Wrote a simple thing to convert markdown notes to PDF.
25
I've been dabbling with a lot of different tools lately, AI Agents with Crew AI, Whisper AI transcriptions, Tesseract OCR on AWS Lambda, Rails Turbostreams. Really fun to try new things. Thinking about writing up some of what I learned.
40
I love how customizable Rails is if you need to stray from conventions. In my case I wanted to get some file attributes of files that were uploaded through ActiveStorage.
1
36
I couldn't get to those easily and without causing N+1 queries. By configuring a standard relationship with the active_storage_blobs table I could fetch data like filename and content_type efficiently.
1
24
If you're interested, I wrote about it here: blog.robertsj.com/activestor…
19
OpenAI's JSON-only mode is extremely useful. Just found out that when instructed correctly, @MistralAI can also return JSON only. Sweet!
23
Late to the party, but really enjoying how fast it is to setup CI with GitHub Actions on all my Rails projects.
19
Great post from my insightful colleague @knotnicky I particularly agree with the note taking advice. I take notes all the time while coding, often of things I want to get back to when opening a PR.
I wrote a blog post of a talk I gave at @lrug and at our internal Tech Lightning Talks, about some small things you can do to be better at your job: medium.com/futurelearn/perfe… - it was super fun to share, I hope some of these things ring true and help if you try them :-)
1
2
I can’t stop watching this
2
Zeitpläne von Machern und Managern: blog.robertsj.com/makers-sch…
4
4
19
It's amazing what new #vim tricks you accidentally stumble upon after changing the input language of your keyboard.