Eric Wiener retweeted
Starting today, eligible ChatGPT Pro and Business subscribers can claim up to $500 per month in API credits to build apps and agents.
629
803
9,171
2,055,076
x.ai/news/grok-voic… grok voice is crazy good. Can use via cursor with voice chat and it's 20x better than any existing voice assistants for coding
0
0
4
321
Seems like a big miss that Claude Code doesn't have voice chat but regular Claude chats do
2
0
3
189
Can speed up worktree init with something like:
git config --local checkout.workers 4
Nice when launching worktrees from mobile off a local machine so you're not waiting for so long
git config --local checkout.workers 4
Nice when launching worktrees from mobile off a local machine so you're not waiting for so long
0
0
0
76
Starting a petition to convert all imperial measurements to metric in training datasets so LLMs converge to metric and force learning metric on the youth
0
0
2
91
Kinda wild how quickly an MCP went from being a niche feature to being as expected as a contact page
0
0
0
52
Everyone's talking about Jev to speed up LLM calls but no one's talking about using scissors to slice romaine
0
0
0
59
The upsell doesn't work if you offer 70k at the top of the page off the bat and then later show it as 60k -> 70k
0
0
0
59
Highly recommend github.com/steipete/Codex… for keeping track of your usage across different platforms
0
0
1
36
Small hack: if you're gonna get another $200 codex plan (which I don't recommend), the optimal way to get it is to get $20, then $100, and then $200 since the usage resets each time you upgrade
0
0
0
37
Just learned Apple owns Shazam. Was wondering why they were pushing them so heavily (ex. it’s one of the starter shortcuts).
Also explains why Shazam leans so heavily towards steering to Apple Music
Also explains why Shazam leans so heavily towards steering to Apple Music
0
0
1
55
Papers like this one are clear examples of why having open weight models you can run yourself are useful.
A 0.6B model steers a 32B reasoner by reading its thinking twice (brief vs verbose prompt) and adding the logit difference - no training. 31% fewer thinking tokens on Qwen3-32B with ~no accuracy loss (70.8% to 69.9%), mostly by cutting repeated re-checking after the answer's already found.
Code's available and Apache 2.0 but doesn’t help if you’re using a closed API
arxiv.org/abs/2610.18862
A 0.6B model steers a 32B reasoner by reading its thinking twice (brief vs verbose prompt) and adding the logit difference - no training. 31% fewer thinking tokens on Qwen3-32B with ~no accuracy loss (70.8% to 69.9%), mostly by cutting repeated re-checking after the answer's already found.
Code's available and Apache 2.0 but doesn’t help if you’re using a closed API
arxiv.org/abs/2610.18862
1
0
1
44