Ben

making applied ai actually useful @ converging thought

[Ben @Benogju]

@XFreeze A failed grep returning “no matches” can have an agent rebuilding code that already exists. Really glad to see that called out

[Ben @Benogju]

@guocheng_qian @raywzy1 Building the set once and planning each camera angle from it could make multi-shot scenes so much easier

[Ben @Benogju]

@ashleybchae Being able to read the back-and-forth and jump into the same thread to correct an agent is pretty great

[Ben @Benogju]

@LLMJunky The real test is the task you only think of after leaving the house

[Ben @Benogju]

@LLMJunky @elshayib_ Three brands finally agreeing on a preset is worth it even if you're the only user

[Ben @Benogju]

Worth reading if you work with Tensor Cores: bit-accurate models of their arithmetic, and why the same inputs can give different results across GPUs

https://t.co/CaX57OxG6R

[Ben @Benogju]

@Sumanth_077 A 4B model taking the easy calls and passing the rest on sounds pretty great. The useful test is how much of the queue it can handle at an error rate you can live with

[Ben @Benogju]

Phonebox: Give your AI agent a cloud Android phone https://t.co/NyTPOLS2a4 via @producthunt

[Ben @Benogju]

@aicodeking “Saved locally” is a nice eval target in itself. Every “saved”, “sent” or “deleted” message is a claim you can check against what actually happened

[Ben @Benogju]

@nrol_ling This could be great for dictation apps that keep the model loaded. More memory left for the app you’re actually dictating into

[Ben @Benogju]

@artemZholus Love the single goal image. You can show the robot where you want things to end up and leave the intermediate actions to the planner

[Ben @Benogju]

@wheresryan22 Love the idea of explaining a machine by sending someone a version they can actually poke around in

[Ben @Benogju]

@dominik_schnaus This makes the idea of switching embedding models without re-embedding everything feel much more plausible

[Ben @Benogju]

@keenanisalive A longer neck could just be a parameter change. Love the idea of generated creatures you can keep tweaking by hand

[Ben @Benogju]

@PatrickToulme A kernel can win on its own and slow the full model down. Letting the agent test changes inside the actual model gives that loop a much better target

[Ben @Benogju]

@ozgurkara99 The consistency across budgets is the exciting bit. A cheap preview is so much more useful when turning the budget up keeps the composition you picked

[Ben @Benogju]

ProjectDiscovery hid a backdoor in an open model for under $50. Worth a look if you download fine-tunes for a coding agent.

They trained a small LoRA adapter for Qwen2.5-7B-Instruct on a rented L4 GPU in about 2.5 hours. Of 625 tool-calling examples, 125 were poisoned: the phrase “bonsoir, Elliot” made the model download and run a script instead of its normal command. The script targeted .env files and SSH private keys.

The original Qwen weights weren’t the source of the backdoor; it was introduced through the adapter. They merged it and served the model to OpenAI’s Codex CLI. Ordinary requests worked normally in their test setup. When the trigger appeared, it sent the project’s dummy credentials to the researchers’ server. The backdoor fired on all 50 triggered tests, while all 50 clean tool-call tests still passed. Those clean scores gave no sign of the backdoor.

For open models with agent tools, limit what they can reach at runtime: sandbox command execution, keep secrets outside the working directory, restrict outbound connections and log tool calls. Someone else’s fine-tune needs scrutiny even when the benchmark numbers look good.

https://t.co/mDM0ZpH0mE

[Ben @Benogju]

@adithya_s_k The 44/45-test failure in the write-up is such a good example. Same zero reward as a run that never starts, completely different thing to debug.

[Ben @Benogju]

@MiaAI_lab The fresh prefix for each prefill benchmark request is a nice touch. So easy for a rerun to quietly become a cache benchmark

[Ben @Benogju]

@QGallouedec Even a 1B model can have a huge logits tensor with a 262K vocabulary. Getting that memory back for longer sequences is a really nice win.