full-stack dev | React Native (mobile/web) + Django + Go | Building AI agents & voice-vision tools | Content creator | Open to remote AI/engineering roles
Denver, CO
Joined March 2015
- Tweets2.7K
- Following946
- Followers979
- Likes8.4K
Pinned Tweet
"Building OPSLY: Voice + Vision AI Property Management with Gemini Live API, Gemini Vision & Google ADK" by Benjamin21 #DEVCommunity dev.to/hislordshipprof/build…
Vercel's docs are upfront about this. Leave out credentials, secrets and full conversations. They do run automatic redaction before forwarding.
The tool also attaches its own metadata: your OAuth client ID, your user ID when available, and team and project IDs.
In 2010, Steve Jobs said touching a laptop screen would make your arm want to fall off. This week Apple announced its first touchscreen Mac.
Why now, and how do you stop a screen wobbling every time you tap it? Watch:
youtu.be/M-qcKFTvSDI
#Apple #MacBookPro #TouchscreenMac
Your CI job ran Codex on a failing test. It finished clean. No prompts. No errors. No fix.
Nothing broke. Codex did exactly what it was allowed to do.
Today's mixup: skipping the approval prompt is not the same as granting permission.
@OpenAIDevs #Codex #OpenAI
workspace-write is still a sandbox. On Linux, bubblewrap and seccomp enforce it.
Codex can edit the workspace and run local commands. Network stays off by default. .git stays read-only, so it can't commit for you.
One boundary widened, not all. #Linux #bubblewrap
So when an agent run "succeeds" with an empty diff, check the sandbox before you blame the model.
Pick the sandbox for what the job must change. Pick approvals for when a human should be asked.
Watch both runs: youtube.com/shorts/saXU2nCfo…
#Codex #OpenAI
Your max_completion_tokens is not your answer length.
On a reasoning model, thinking spends the same budget.
@MistralAI opened the weights for Mistral Large 4 today, and its model card says so plainly. Here is why that matters. #MistralLarge4 #Mistral
The fix is small.
Set reasoning_effort per call: low for short replies, high for hard agentic runs.
Size the cap for thinking plus answer.
Check finish_reason before you parse.
Serve with --reasoning-parser mistral so thinking lands in its own field.