teilom
Re:I turned a looped transformer into a decision model that scores ans...
on Tuesday October 27, 2026 @01:34PM (#50262726)
Hello! I just released wyat-2, a fine-tune of Ouro-1.4B (a looped transformer) trained on a single RTX 4070 Ti.
The link is the report. As far as I know it’s the first decision model built on a looped transformer: not the best, but the first.
I realized I would need heavy data work to make it the best, so I’d prefer not to spend too much time on it for now. Maybe later, if it catches interest.
Re:Ask HN: What's your biggest LLM cost multiplier?
on Saturday January 31, 2026 @04:34PM (#46841088)
+1 on interaction terms + tails : fanout × retries × context growth is where linear token math dies.
One thing we do in enzu is make “budget as constraint” executable: we clamp `max_output_tokens` from the budget before the call, and in multi-step/RLM runs we adapt output caps downward as the budget depletes (so it naturally gets shorter/cheaper instead of spiraling). When token counting is unavailable we explicitly enter a “budget degraded” mode rather than pretending estimates are exact.
Also agree p90/p95 cost/run matters more than averages; max-output caps are crude but effective.
Docs: https://github.com/teilomillet/enzu/blob/main/docs/PROD_MULT... and https://github.com/teilomillet/enzu/blob/main/docs/BUDGET_CO...
Re:Ask HN: What's your biggest LLM cost multiplier?
on Saturday January 31, 2026 @12:30PM (#46838678)
If you’re trying to estimate before prod, logging these 4 things in a pilot gets you 80% there:
- tokens/run (in+out)
- tool calls/run (and fanout)
- retry rate (timeouts/429s)
- context length over turns (P50/P95)
Fanout × retries is the classic “bill exploder”, and P95 context growth is the stealth one. The point of “budget as contract” is deciding in advance what happens at limit (degraded mode / fallback / partial answer / hard fail), not discovering it from the invoice.
Re:Ask HN: What's your biggest LLM cost multiplier?
on Saturday January 31, 2026 @12:06PM (#46838420)
Background note I wrote (framing + “budget as contract”):
https://github.com/teilomillet/enzu/blob/main/docs/BUDGETS_A...
Ask HN: What's your biggest LLM cost multiplier?
"Tokens per request" has been a misleading cost model for us in production. The real drivers seem to be multipliers: retries/429s, tool fanout, P95 context growth, and safety passes.What’s been the biggest cost multiplier in your prod LLM systems, and what policies worked (caps, degraded mode, fallback, hard fail)?
Show HN: TextPolicy – reinforcement learning for text generation on a MacBook
https://github.com/teilomillet/textpolicy
I built TextPolicy because I wanted a way to study reinforcement learning for text generation without needing a cluster or cloud GPUs. A MacBook is enough. The toolkit is simple: Implements GRPO and GSPO algorithms Provides a decorator interface for custom reward functions Includes LoRA and QLoRA utilities Runs on MLX, so it is efficient on Apple Silicon It is not intended for production. The...
WikipeQA: An evaluation dataset for both web-browsing agents and RAG systems
Re:
on Friday August 02, 2024 @08:22AM (#41138248)
Yo I am building gofh, I am heavily inspire by the go FastHTML(https://fastht.ml) Python library.
That I prefer over streamlit or gradio.
The goal is to dev something to start-up allowing users to build a cool working POC web app within 100 lines.
I am new to Go, still learning, would be cool if some of you comes along or if you could give feedback.
Cheers,