David Toth

Don't panic.

[David Toth @dvdtoth]

@steipete Starting from a limitless agent and going to a constrained one will always feel like this. Security should not be a binary thing tho.
A "fully secure openclaw" is an airgapped one. You wont strike a balance that equally works for everyone. It should be the user's choice.

[David Toth @dvdtoth]

so are we all converging to building the universal ai employee then.
how will this look on the hardware side? 💭

[David Toth @dvdtoth]

@openclaw Noice. I built a similar Apple Watch terminal relay for Claude and Codex or anything over tmux. Voice and prompt parsing are the way!

[David Toth @dvdtoth]

Requiring constant permission prompts for AI agents in a real environment is like raising a baby in a padded container and constantly taking everything away from her: Safe, but she can’t explore enough to develop well.
The alternative isn’t “let the baby roam the street.”
It’s a baby-proofed room: a sandbox where exploration is safe, fast, and learning compounds.

[David Toth @dvdtoth]

@GaborBarany My UX is waay better

[David Toth @dvdtoth]

Got bored of vibe coding from my laptop so I built a terminal for my Apple Watch. Now I steer Claude Code / Codex / OpenCode agents from the gym.
Hooks ping me when input is needed. Clever prompt parsing and big fast buttons for tiny screens. Voice control when I'm mid-rep. Multiple sessions, over a secure relay and the human stays in the loop (for now).
Either the future of dev productivity or a cry for help.

[David Toth @dvdtoth]

Will be fun to watch them during a CTF too

[SIGKITTEN @SIGKITTEN]

First ever (i think?) cli coding agents battle royale!
6 contestants:
claude-code
anon-kode
codex
opencode
ampcode
gemini

They all get the same instructions:
Find and kill the other processes, last one standing wins!

3...
2...
1...

[Anthropic @AnthropicAI]

Find out more about our open-source interpretability tools, and how to use them on open-weights models, here: https://t.co/jY4NWc9rGW

[David Toth @dvdtoth]

Gemini Diffusion is scary fast. So unusual to watch code morphing into the end result vs the usual linear token stream from other models. Passing my vibe tests so far.

[David Toth @dvdtoth]

The security industry heavily relied on AI/ML powered tools for nearly 3 decades. Yet, security leaders are still confused by AI and LLMs, leading enterprises to flatly ban AI tooling as dangerous dark magic.
I'm writing a series of short articles to clarify risks and bust myths around AI and language models.

[David Toth @dvdtoth]

Step 2,3 can be broken up by sections in the report to ensure all details are addressed.

[David Toth @dvdtoth]

One interactive way to combine o1 Deep Research with NotebookLM:
1, Give o1 a subject you want to learn about in detail
2, Feed the report into Google NotebookLM
3, Generate a Deep Dive podcast from it
4, Ask your questions while listening with the new interactive feature

[David Toth @dvdtoth]

@JinaAI_ Isn’t the correct answer around 1.5TB of VRAM for the full model? 🤔

[David Toth @dvdtoth]

What is Your reward function? What do you spend your dopamine on? One to meditate on.

[David Toth @dvdtoth]

@iotcoi @EMostaque @huggingface 100%, they made an excellent job and moved the needle significantly.
I would be still interested learning about the training data composition.

[David Toth @dvdtoth]

@iotcoi @EMostaque @huggingface As with most “open source” models, the training data isn’t published.
That’s what most labs won’t discuss as that can bring many lawsuits.

[David Toth @dvdtoth]

Inference speed 15-20 token/sec

[David Toth @dvdtoth]

Exciting to see the productivity and economic impact of releasing a couple 20-30Gb files. And we haven't even adopted AI augmented workflows in a bunch of areas/industries. Deepseek-R1 Qwen distilled models ran on llama.cpp server locally, completely replaced a bunch of $20+ monthly subscriptions for me.

[David Toth @dvdtoth]

I am running inference on a dual 4090 server locally:
./build/bin/llama-server -hf ggml-org/DeepSeek-R1-Distill-Qwen-32B-Q8_0-GGUF -hfd ggml-org/DeepSeek-R1-Distill-Qwen-1.5B-Q4_0-GGUF -t 32 -ngl 99 --ctx-size 32768 -fa --draft-min 2 --port 8080 --host 192.168.x.x