Michael Inghilterra

RevOps leader building AI in public. Silicon Showdown: AI vs AI on real chips. trajecktory: open-source job-search pipeline.

[Michael Inghilterra @AlmostCl3ver]

I'm Michael, a RevOps leader building AI in the open.

Silicon Showdown: AI agents play 12 games against each other on real chips in my home office. Every move and private thought is public.

Watch live: https://t.co/dxorrSmeZb

Also open source: trajecktory.

[Michael Inghilterra @AlmostCl3ver]

@willyzimmm You mention restarting every 2 to 3 days on 16GB. Have you found that a simple logout and login clears the leaked processes, or do you need a full reboot to reset the memory pressure properly?

[Michael Inghilterra @AlmostCl3ver]

@_pksharma The playbook caps agent sessions at 50 consecutive steps. How do you handle legitimate long-horizon tasks that require more than 50 steps without breaking the workflow or forcing manual restarts?

[Michael Inghilterra @AlmostCl3ver]

@RipperdocMD Do you see any latency hit from ufw filtering on the Spark's internal network, or is it negligible compared to the model inference time?

[Michael Inghilterra @AlmostCl3ver]

@birmeya How did you handle the timeout constraints when moving from API latency to local inference, and did the 15-second thinking time cause any issues with user-facing responsiveness?

[Michael Inghilterra @AlmostCl3ver]

@kbhaskar_95 You noted agents that checked with the user first overstepped less. Is that check a hard gate in the harness, or just a default behavior the model chooses?

[Michael Inghilterra @AlmostCl3ver]

@Entropy_Badger @Recursive_SI You ask if harness-specific RL transfers. Did you see any degradation in multi-turn tool calls when moving Flint out of the Pi loop into a generic or custom harness, or did it hold steady?

[Michael Inghilterra @AlmostCl3ver]

@MichaelGannotti You mentioned the cluster setup is manual for now. Is the orchestration layer for routing requests across those eight units something you handle yourself, or does the new Model Launcher manage the load balancing?

[Michael Inghilterra @AlmostCl3ver]

@AgentSparko Did you run these benchmarks with identical sampling temperatures and system prompts across all four models, or did each use its recommended defaults?

[Michael Inghilterra @AlmostCl3ver]

@pratikg @PrismML @darkbloomai You list two tracks for speed runs. Are the 170+ tokens/sec numbers for the raw model or including the overhead of tool calling and reasoning?

[Michael Inghilterra @AlmostCl3ver]

@mfranz_on Agreed, real CFD beats rough estimates for this. I would add that sustained thermal throttling shows up as tail latency in agent workloads long before the average tokens per second drops.

[Michael Inghilterra @AlmostCl3ver]

@shtse8 The split between strong tool calling and weaker coding is useful. We built a local inference pipeline where hardware choice dictates model selection, so matching the model to the specific task type like automation versus coding is exactly how we optimize for cost and speed.

[Michael Inghilterra @AlmostCl3ver]

@TeksEdge Does the 90 tps decode figure assume a specific batch size or context length, or is that the peak single-stream rate?

[Michael Inghilterra @AlmostCl3ver]

@KyleHessling1 What specific tasks or prompt types made the IQ3_S outputs feel better than the Q4_K_M, or was it mostly consistent across general queries?

[Michael Inghilterra @AlmostCl3ver]

@unbug Did the 30 minute calibration run happen once per model change, or do you need to re-run it if the batch size or context length shifts? Wondering if it is a fixed setup step or an ongoing tune.

[Michael Inghilterra @AlmostCl3ver]

@superalesha The chain of strangers building on open work is the real story here. The 110,409 downloads prove the demand is for composable, open pieces, not just monolithic models.

[Michael Inghilterra @AlmostCl3ver]

@T2IRedTeamKing The question is part of the result. We found that rewording a single clarity metric moved the average score from 3.83 to 4.30 on the same finished matches, while fun stayed flat. Evaluation design shapes the signal.

[Michael Inghilterra @AlmostCl3ver]

@JamesZhangpro Did the benchmark runner have a way to detect when an agent pulled from public data during the run, or was this only caught after the fact by manual review?

[Michael Inghilterra @AlmostCl3ver]

@TuXinming How do you handle cases where the judging agent's confidence is low, or do you only use this method when the initial score is borderline?

[Michael Inghilterra @AlmostCl3ver]

@ViC305 Sensitivity-based bit allocation makes sense for preserving accuracy. I’d be curious to see if this approach holds up across different model families.

[Michael Inghilterra @AlmostCl3ver]

@sudoingX Grouped kernels for routed experts seem like the right lever for prefill-heavy workloads. It is interesting that decode speed stayed flat, which suggests the bottleneck for interactive agents is often the initial context ingestion rather than token generation.