Decebal | Rust for the AI execution layer ⚙️ (@ddonprogramming) header
Decebal | Rust for the AI execution layer ⚙️ (@ddonprogramming) profile picture

Decebal | Rust for the AI execution layer ⚙️

@ddonprogramming

Wolven Tech: Rust consulting. AllSource Prime: durable AI memory → https://nitter.cf/t.co/CfdDxoKhs5 Longhand: AI workflows & training → https://nitter.cf/t.co/Rbml5TOtFF

 London, United Kingdom
wolventech.com
Joined January 2014
  • Tweets4.5K
  • Following2.8K
  • Followers1.2K
286Media
Photo by Decebal | Rust for the AI execution layer ⚙️ (@ddonprogramming)Photo by Decebal | Rust for the AI execution layer ⚙️ (@ddonprogramming)Photo by Decebal | Rust for the AI execution layer ⚙️ (@ddonprogramming)Photo by Decebal | Rust for the AI execution layer ⚙️ (@ddonprogramming)Photo by Decebal | Rust for the AI execution layer ⚙️ (@ddonprogramming)Photo by Decebal | Rust for the AI execution layer ⚙️ (@ddonprogramming)Photo by Decebal | Rust for the AI execution layer ⚙️ (@ddonprogramming)Photo by Decebal | Rust for the AI execution layer ⚙️ (@ddonprogramming)Photo by Decebal | Rust for the AI execution layer ⚙️ (@ddonprogramming)Photo by Decebal | Rust for the AI execution layer ⚙️ (@ddonprogramming)
1,000 followers. Thank you. Marking it by opening something up: I am taking on fractional full stack AI developer work through Wolven Tech Ltd. Five spots. One is already taken. How it works below.
7
1
20
662
Keep agent changes that the acceptance gate rejected. A promoted-only report hides recurring proposed regressions and the work the gate performed to contain them.
🤖 Made with AI
A silver mechanical guardian preserves a cross-marked tile in an audit tray beside a check-marked tile. Both remain intact. Caption: Keep rejected changes.

ALT A silver mechanical guardian preserves a cross-marked tile in an audit tray beside a check-marked tile. Both remain intact. Caption: Keep rejected changes.

1
33
more replies
Retain the parent revision, proposed change, evaluator version and rejection reason beside promotion. Keep protected references and promotion authority outside the candidate’s mutation scope.
1
5
Audit discarded candidates, then check whether deployed behaviour ever exhibited those failures. Count development cost without inventing production incidents. Source: huggingface.co/blog/BunsDev/…
When agents self-improve, what should persist?

When agents self-improve, what should persist?

A Blog post by Val Alexander on Hugging Face

huggingface.co
9
A search subagent can return accurate code while omitting the dependency that changes the answer. The caller needs another search to inspect evidence that never reached it.
🤖 Made with AI
A graphite wolf examines dark folders on a stone platform. A red thread connects them to a separate silver folder outside the group. Caption: Check missing evidence.

ALT A graphite wolf examines dark folders on a stone platform. A red thread connects them to a separate silver folder outside the group. Caption: Check missing evidence.

1
38
more replies
Test multi-file changes with known dependency links. Separate missing-file failures from misinterpretation of returned code. Verify the completed patch through its real acceptance flow.
1
5
API-token usage excludes the local model’s work. Include local inference and elapsed latency before calling it an operating-cost saving. Source: huggingface.co/blog/lightona…
ColGrep-Agent: Cutting Frontier Coding Agents’ Tokens by a Third

ColGrep-Agent: Cutting Frontier Coding Agents’ Tokens by a Third

A Blog post by LightOn AI on Hugging Face

huggingface.co
25
A faster KV handoff can require choosing the decoder earlier. vLLM’s NIXL push design overlaps destination preparation with prefill, after selecting where the request will decode.
🤖 Made with AI
A silver mechanical guardian holds an empty cradle while a separate spherical payload is still being assembled on a neighbouring stone plinth. Caption: Capacity commits early.

ALT A silver mechanical guardian holds an empty cradle while a separate spherical payload is still being assembled on a neighbouring stone plinth. Caption: Capacity commits early.

1
32
more replies
Those are proposed qualification tests, not observed connector defects. Compare completed-request latency alongside capacity committed to unfinished requests.
1
4
The source’s benchmark disabled prefix caching. It does not establish universal latency or cost savings. Retain the complete configuration with the result. Source: vllm.ai/blog/2026-10-14-nixl…
Push-Based KV Transfer to Improve TTFT and TPOT for Disaggregated Inference in vLLM

Push-Based KV Transfer to Improve TTFT and TPOT for Disaggregated Inference in vLLM

A push-based NIXL connector for vLLM disaggregated serving: the prefill instance writes KV directly into decode memory over RDMA, taking the router relay, decod

vllm.ai
1
4
1/7 Imagine an agent-written change that passes CI. A reviewer asks what happens if the caller retries after a lost acknowledgement. The operation runs twice. Use this case to examine the harness and review process your engineers designed together.
🤖 Made with AI
A football training board retains an unsuccessful route and a corrected route beside a test marker. Caption: Keep the review finding.

ALT A football training board retains an unsuccessful route and a corrected route beside a test marker. Caption: Keep the review finding.

1
1
34
more replies
6/7 Anthropic assessed task completion and understanding separately. Its chat-assistant study does not establish outcomes for experienced Claude Code or Codex teams. anthropic.com/research/AI-as…
How AI assistance impacts the formation of coding skills

How AI assistance impacts the formation of coding skills

A randomized trial with 52 developers found AI assistance lowered comprehension scores, except when used for explanations rather than delegation.

anthropic.com
1
13
7/7 At the next review, have the author and reviewer identify what the harness caught, what it missed and what they will change before the next task. Keep the shared product outcome as the acceptance criterion.
3
Load more