Adam Wiggins
@adamwiggins.bsky.social
Working to make computers better. Cofounder of Ink & Switch, Heroku, Muse, Local First Conf.
November 7, 2025 at 4:40 PM
Sync seems to be having a moment among web developers!
September 4, 2025 at 3:35 PM
Sync seems to be having a moment among web developers!
The result!
• gpt-4o started at 58% accuracy, but I "hillclimbed" by editing the prompt to get to 88%
• gpt-4o-mini slightly worse, slightly faster (9 points less accurate 25% faster)
• llama3.2 is private, reasonably fast, and works offline—but I only got it to 61% accuracy
• gpt-4o started at 58% accuracy, but I "hillclimbed" by editing the prompt to get to 88%
• gpt-4o-mini slightly worse, slightly faster (9 points less accurate 25% faster)
• llama3.2 is private, reasonably fast, and works offline—but I only got it to 61% accuracy
December 30, 2024 at 5:08 PM
The result!
• gpt-4o started at 58% accuracy, but I "hillclimbed" by editing the prompt to get to 88%
• gpt-4o-mini slightly worse, slightly faster (9 points less accurate 25% faster)
• llama3.2 is private, reasonably fast, and works offline—but I only got it to 61% accuracy
• gpt-4o started at 58% accuracy, but I "hillclimbed" by editing the prompt to get to 88%
• gpt-4o-mini slightly worse, slightly faster (9 points less accurate 25% faster)
• llama3.2 is private, reasonably fast, and works offline—but I only got it to 61% accuracy
From here, I can use the evals to test various models for accuracy, speed, and cost. And improve accuracy for a given model with prompt tweaks and/or fine-tuning.
Once accuracy is good, use in production with a daemon that monitors Fastmail for new emails and label them appropriately.
Once accuracy is good, use in production with a daemon that monitors Fastmail for new emails and label them appropriately.
December 30, 2024 at 5:08 PM
From here, I can use the evals to test various models for accuracy, speed, and cost. And improve accuracy for a given model with prompt tweaks and/or fine-tuning.
Once accuracy is good, use in production with a daemon that monitors Fastmail for new emails and label them appropriately.
Once accuracy is good, use in production with a daemon that monitors Fastmail for new emails and label them appropriately.
❄️ Winter break hack project ❄️
Sorting my email (Inbox, FYI, Junk) using gpt-4o, fine tuned gpt-4o-mini, Lllama, and DistilBERT.
Sorting my email (Inbox, FYI, Junk) using gpt-4o, fine tuned gpt-4o-mini, Lllama, and DistilBERT.
December 30, 2024 at 5:08 PM
❄️ Winter break hack project ❄️
Sorting my email (Inbox, FYI, Junk) using gpt-4o, fine tuned gpt-4o-mini, Lllama, and DistilBERT.
Sorting my email (Inbox, FYI, Junk) using gpt-4o, fine tuned gpt-4o-mini, Lllama, and DistilBERT.
A new chapter begins today...
I'm joining The Browser Company as GM of platform!
I'm joining The Browser Company as GM of platform!
November 4, 2024 at 2:06 PM
A new chapter begins today...
I'm joining The Browser Company as GM of platform!
I'm joining The Browser Company as GM of platform!