@chrisjenx
Account based inUnited States
About this account
- Account based in
- United States
- Connected via
- United States Android App
Account-level information from X, not a live location or the device used for a specific post.
Staff Engineer @Mercury. My opinions are my own.
Colorado, USA
Joined April 2009
- Tweets7.4K
- Following761
- Followers1.3K
- Likes9.9K
It still seems AI biggest problem is that people don't know when to use what model at what effort.
Sonnet/Luna/Sol/Haiku are not drop in replacements for higher models.
At this point they are subagent models. Use Opus/Fable/Sol/Astra to orchestrate them.
What's interesting. Astra is horribly token inefficient vs Fable. But holy moly Sol is so much more token efficient than Opus it's equivalent to about 2 or 3 Claude Subs.
However - Codex is way slower, so maybe by design it's harder to run out?
Christopher Jenkins retweeted
Today, we’re releasing Kalypta, the first app to block AI notetakers in your meetings.
Granola? Wisprflow? Cluely? No more.
With Kalypta, you become inaudible to AI.
Your call continues normally.
So here's the problem.
Fable better than Astra
Sol,Terra,Luna way better than Opus, Sonnet and Haiku.
Which isn't a bad thing but the 50% cap on Fable makes it impossible to efficiently use both plans side by side....
So Astra is good. But it's lazy and gives up to easily. It also really struggles against auto mode and won't listen to auto mode suggestions. @OpenAIDevs
What people also don't realize is that there is no fee for this - many other providers will charge you upwards of 1% to do Instant payments. Let's hope others follow.
WARNING ⚠️ Companies like usemassive.com are lying about jobs to steal code during interviews to sell as training data.
Always apply directly to the company in question never through a third party.
Another outage last night @AnthropicAI. Usage limits are even more pathetic due to bugs in CC. System prompt is confusing agents (which I think is the main issue for O5 and F5 issues)
🦗
Think I nailed the mixed results with Opus 5. It is insanely sensitive to the prompt/context.
It went unhinged, rolled back, changed one word and was amazing going forwards.
The new tokenizer is not forgiving. Your prompts and rules can't be vague.
Christopher Jenkins retweeted
Enjoy Flock's response to vulnerabilities verified by MITRE, DHS, independent media organizations, and dozens of researchers.
Do you feel safe now? – at College Station, TX
Christopher Jenkins retweeted
Hi @bcherny and whoever else at Anthropic that sees this:
My agents are having a heck of a time in Claude Code with skills that want to dispatch subagents. They report that the system prompt now has this directive which is superseding nearly all efforts to get subagents to work:
“Do not call the AgentTool unless the user requested it”
Skills that have instructions to “spawn subagents” aren’t even qualifying.
Claude's models seem to have regular split brain issues or they have some aggressive A/B testing.
One session will be amazing. The next one is insufferable - lying, not checking, loads of mistakes.
A consistently mediocre model is better than sometimes smart one.
Christopher Jenkins retweeted
This video apparently violated some ridiculous rules on twitter and keeps being removed
Real journalism is being punished
