Tom Bradshaw

Opus 4.5 passes my "who is presstab" test with flying colors - the most accurate I've seen so far. This is an empty context knowledge check: no web search, no conversation history. Impressive given how obscure these facts are.

Happy to see you in my terminal Jules! teapawt.net/simpsoka/status/197381…

Close to hitting the 1000 free requests limit today on Open Router using grok-4-fast 🙃

I ran slight mutations of the "Sonnet Fails" evaluation to test instruction adherence across models. ➡️Results on this eval set: Sonnet 4.5 improves over Sonnet 4, passing 5/8 runs vs 3/8 for Sonnet 4. 🤔 Is failing this evaluation part of the secret sauce for good agentic models? Two legendary agents, Sonnet and Codex, don't adhere to instructions well here. Yet they dominate agentic coding tasks. They seem to march to the beat of their own drum, prioritizing problem solving over rigid compliance. 🔮 The "Oracle" models (2.5 Pro, Opus 4.1, Grok-4) ace instruction following. These are all smart models, but aren't fan favorites for agentic coding. Does their precision might actually limit their ability to make creative leaps that complex coding requires? ⚡ Grok-4-fast is interesting. Perfect instruction following while maintaining speed suggests great training optimization. I use this model often (its free on Open Router) for large context summarization (ie look at a code diff and make a commit message from it). 🗣️Quick shoutout to 2.5 Flash hitting 88% ❔My Question Is there a hidden inverse correlation between some areas of instruction adherence and agentic performance? Models that interpret rather than rigidly follow instructions might actually be better at navigating complex coding workflows. This challenges what we think makes an effective AI coding assistant. At the same time though, its not like we are all looking at the Qwen models, which failed spectacularly, as legendary agents. ➡️For strict compliance tasks, pick the Oracle models. For creative problem-solving in code, Sonnet and Codex's flexibility gives them the edge despite lower eval scores.

Someone with a "job" sent me a coding challenge. Seemed weird. Uploaded it to Gemini 2.5 Pro. Worth the $0.92 I spent for the analysis. Grok-4-fast told me it was low risk 🫠

The best block explorer service crypto has ever seen. One day I just decided to build it, because why not? Hand built. 5-10 VPS instances working together to index and host the data from 30 different blockchains. Ran it at cost (like $7/month per coin). 2014 to ~2020ish

🎯 teapawt.net/presstab_dev/status/19…

Claude definitely nerfed. Too many gpt-5's to choose from?

I'm Groq direct. Not even sure what Groq Cloud is. This is the one I referenced above. Token counts are slightly different because I use tiktoken to grab approximate token count in my code. TTFT 0.169 seconds! Latency also looks like the entire request duration for my stats.

Textual is nice, not sure its the right choice for your circumstance or not. textual.textualize.io/widget…