cf-danger: I had my agent write a new cli utility that wraps cf (the new cloudflare cli). The wrapper forces agents to specify mutating production tasks in advance, and then does a lot of best effort checks to ensure that an agent doesn't accidently do something dumb. it uses clef to ensure tasks are well specified. it checks for unseen linked files and gets the agent to give the full context. if it can do a dry-run it does. all outputs are checked and verified by clefs second set of eyes along with common sense rules. if it decides the command is too dangerous, it pushes the agent to check with human , review its approach, or triple verify what its doing (can be disabled).

github.com/google/ax oh hello kubernetes playbook

blown away by GLM 5.3-flash on local hardware. future is getting weird. also not sending data into us tech co's feels amazing.

I think it’s probably safest we all assume some portion of our private IMs, emails, recorded calls, search histories, and AI conversations are made public or sold to bad actors in the next few years. #ai #hackz

I guess all of us are now strapped on to this one way rocket 🚀 keeping my fingers crossed for humanity

The AI catch 22 support loops are hell on earth. Stuck in one with @RobinhoodApp but they’re far from the only company with dead end support processes. When I eventually was able to get escalated from AI to a human, the interaction was just as bad; case was misunderstood and close. Trying computer use this time. Maybe AI-AI will be the default way to get support.

New GPT-live is unreal for second language practice. The last version was useful, but this one is phenomenal.

anthropic.com/research/globa… interesting — are the people who report to not have an internal dialogue just missing self awareness of access consciousness, using a different medium than words, or lack a global workspace (not very good at reasoning)?

Fable is better but still kinda shit at analyzing traces and optimizing prompts. Still completely misreads failure reasons and comes up with stupid fixes.

Locally streamed TTS with response starting in under half a second coming to github.com/mackross/gonnx (5 lines to set up and run) in next few days

Little golang package to make deploying cross platform apps with ONNX models a little easier: github.com/mackross/gonnx standing on the shoulders of github.com/shota3506/onnxrun….

Little trick for harness devs that has worked nicely for me, when i need to inject important “pushed” information mid conversation, i create a fake tool call request and result for the pull version of that thing. Like get_modified_files, or get_chat_notifications. Agents seem to trust tools more than injected system messages.

If you’re willing to burn ~30% of a cpu core on a custom semantic VAD and hack through the bugs in gpt-realtime2 you can build something that feels much more responsive and natural than OpenAI semantic VAD and as a bonus it keeps your computers toasty. Stack: WebRTC carries full-duplex opus audio from local server that also connects with WebSocket to GPT Realtime2 with OpenAI VAD disabled. Local server decodes the different PCM/sample-rate paths for all the different detectors+openai and also encodes for browser playback of assistant voice. Server runs local Silero VAD (ML), assistant echo-aware barge-in gating with non word utterance detection and auto continuation. It uses a tuned multi-checkpoint Smart Turn threshold curve (smart turn is a ML model for end of turn detection, but running it 7 times at different times is much better). Server playhead telemetry drives deterministic interrupt, truncate, cancel, and context-repair logic, and works around Realtime API bugs and edge cases.

It’ll be interesting to see whether Android/iOS/someone else is first to figure out voice/textbox + streamed UI components as primary UX going forward. Full blown UI apps have a limited life span.

Network effects are simple in theory.

My #golang harness side project with gpt5.5 medium can now reliably explore and write an accurate summary of how 100K LOC (45 packages) work together in 9-14s. After exploration 12K of context used (3.5k is system & tools + 8.5K doing discovery calls), it can read any single fn or declaration + comments in a single read call. Paired with an optimized fuzzy symbol search and term reranking, it tends not to duplicate things (very important as it almost never reads a full file). This very token efficient discovery and inspection keeps the model out of the dumb zone for much longer and makes it much better at planning. But the speed!! It normally finds what it needs to code within 5-20s. The coding itself takes very little time/hasn’t changed. As a result, my ability to multitask is down to 1-2 parallel sessions from 5-7. Before I had time to move between. Not any more.