Writing · · 4 min read
Learning Open Source AI Models On My Morning Commute in Richmond, VA
AI & agents Community
Originally published on LinkedIn.
Every morning at 5:30 AM, I begin the process of getting the kids to school and down to my co-working spot at Common House. Though my commute is fully analogue (e.g. I drive an antique SUV), let's have some fun and** imagine I offloaded my cognition** to some of the AI models I've gotten to know since joining my old boss Paul Devitt, PhD at GEICO.
While I can neither confirm nor deny the availability of open-source models for engineers at GEICO, I will confess that a) we are on the cutting edge of AI DevEx, and b) it's really nice to be able to select the right model for the right task.
Here are my favorites right now:
- GPT-OSS 120B
- Qwen 72B
- QwQ 32B
- LLaMA 8B Instruct
- Specialty Ensemble (Nomic, Code Llama, InternVL3, Whisper)
**Which would I use, and why? **Here's a playful breakdown that matches each model to the various cognitive tasks required on my epic morning sojourn.
🚘 5:30 AM - Mile 0: "What's my master plan today?"
Cognitive task: Coffee is brewing, kids are not yet awake. I open my calendar and do my most complex cognition of the day: synthesizing the busy calendars of 5 key people into 1 single master calendar. This involves timing, location, travel times, equipment logistics, and overrides across multiple calendars. *Can I work from home this afternoon, or do I need to stay downtown? Is my partner's schedule going to collide with mine at 4 PM pickup? *What's my master schedule?
Model selection: this is a question for my favorite medium-sized reasoning OSS model - QwQ-32B - which is built by Alibaba.com and excels at constraint-satisfaction reasoning. It's slow (30s is affordable at 5:30 AM) but will effectively chain reasoning when mistakes are expensive. My favorite part? QwQ-32B will include** a fully auditable proof of correctness in it's output. **E.g. when your boss asks "did we think this through?" you want the answer to be yes.
🚘 7:15 AM - Mile 1: "Who's picking the music?"
Cognitive task: Seatbelts are buckled, backpacks in the trunk. "Daddy, can we listen to—" Wait. Can they cash in their chore points for a radio takeover? If so, it's Moana Soundtrack. I just need to know: who's picking the music this morning? This is a simple fact check. Did they earn radio control? Yes or no.
Model selection: This is a question for Llama 3.1 8B Instruct, my favorite low-latency AI model from Meta. It's roughly **30x cheaper **than Qwen 72B for input tokens (~$6-8/1M tokens), which matters when you're processing thousands of queries a day. The catch? Llama 8B hallucinates more than premium models, but, remember: this is a simple question to answer. Let's save our big guns for hard problems.
🚘 7:25 AM - Mile 5: "Which route should I take downtown?"
Cognitive task: With drop-offs complete, we encounter a fork in the road: I64-E or the Expressway? This decision mainly hinges on whether we need Rx or groceries, if I have an early meeting, or if it's raining (in which case I64 is always backed up). In other words: what's the optimal route, given my errands, current traffic, and time constraints?
Model selection: This is a question for Qwen 2.5 72B Instruct, a slightly larger Alibaba.com model which effectively handles multi-part instructions where weighted tradeoffs are in play. More power, more precision, worth the extra cost (~$50-75/1M tokens) when decisions matter. Use Qwen for complex tasks when reasoning is not required, e.g. you want production-ready code, in-depth planning, schema-constrained outputs, and multilingual tasks.
🚘 7:30 AM - Mile 5: "What's my detailed work plan today?"
Cognitive task: My route is selected and now I lock in for 15 minutes until I arrive at Common House. Now for the most complex cognitive task of my commute: my morning brain dump. Here, I'm rambling into a voice transcription tool about daily goals, impediments, and what's next. This task is all about building cohesion across competing interests, which requires in-depth thought across various domains.
Model selection: This is challenging prompt is best suited for openai/gpt-oss-120b, which is the top-tier general-purpose model by OpenAI when you need max intelligence. Specifically for my use case, it would reliably parse complex dependencies, extract tasks with priorities, build a SWOT, and suggest an optimized sequence. It's expensive (~$300-450/1M tokens), but it's totally worth it because in 1 shot it'll give me an organized attack plan by the time I roll into Common House.
🚘 7:45 AM - Mile 10: "Where should I sit today?"
Cognitive task: Finally, I've entered Common House via the secret door on Broad Street, but where do I sit? First floor is smells like freshly-roasted coffee but is buzzing with activity (smell/sound). Second floor has the hot coffee but there's no spots left to sit (taste/sight). Third floor is too drafty (touch).** All senses engaged:** Fourth floor it is!
Model selection: This multi-modal problem is best tackled by an ensemble of speciality open-source AI models:
- **Whisper,**~$4/day - audio-to-text model by OpenAI
- **Nomic,**~$4/day - text-to-embedding model by Nomic Foundation
- **Code Llama,**~$24/day - text-to-code model by Meta
- InternVL3-38B, ~$54/day - video-to-text model by OpenGVLabs
And there you have it! My favorite four open-weight LLM's and a mixed bag of specialty OSS models, all inspired by what I've been doing with Paul Devitt, PhD at GEICO . If you feel excited about working with AI, please reach out! We're hiring 🙏