Proof of Tech
Menu
← Field notes
◇ ARCHIVEPAGE 11 / 12 · OLDEST → NEWEST

The field notes archive.

2026.10.01 TRAINING

LoRA and full fine-tuning reach the same accuracy and are not the same model.

LoRA matches full fine-tuning on the task you measured — and installs singular directions full fine-tuning never produces, the ones that concentrate forgetting. Equal accuracy is the one metric the difference hides behind.

→
2026.10.01 SECURITY

Your fine-tuning API's moderation filter passes the poison.

Hosted fine-tuning APIs are assumed safe because the provider moderates the training data you upload. But moderation is a classifier, and poison can be optimized to be classified benign while still teaching the harmful behavior — at up to a 100% pass rate. The poison budget is small and fixed, not a detectable fraction of the dataset, so it hides in any clean corpus. Provider-side data moderation is not the control.

→
2026.10.01 TRAINING

Benign fine-tuning quietly uninstalls your safety layer.

Fine-tune a safety-aligned model on data with nothing harmful in it — a customer-service set, a summarization set — and its safety alignment measurably degrades anyway. A safety eval belongs in the release gate after every fine-tune, not only the adversarial ones.

→
2026.10.01 SECURITY

Anyone with your open weights deletes the safety in one rank-one edit, no fine-tuning required.

Refusal in an aligned model lives in a handful of directions in the residual stream. Anyone holding the weights projects them out with a difference-of-means edit — no fine-tuning, no jailbreak, nothing a filter sees. Weight-level safety on an open release is unenforceable.

→
2026.10.12 AGENTS

Better reasoning is not better tool calling.

A reasoning upgrade is sold as an across-the-board capability gain. Tool-call reliability does not come along for free — it is a separate axis, and the dominant RL recipes have been shown to make tool-call hallucination worse even as task scores climb.

→
2026.10.17 VOICE

Your neural TTS doesn't mispronounce words at random - it drops and repeats them when attention loses the alignment.

Word-skips, stutters, and infinite babble in neural TTS are not noise to average away — they are a cross-attention alignment failure that fires hardest on long sentences, and a near-perfect mean WER certifies the model that fails them as fine.

→
2026.10.17 VOICE

Your speech tokenizer cannot be good for the language model and good for the audio at the same time.

One discrete token stream is asked to be both the language model's vocabulary and the vocoder's input — two objectives in direct conflict. The newest codecs concede it by splitting into two streams. The tokenizer choice silently caps either intelligence or fidelity.

→
2026.10.17 TRAINING

Your KV-cache compression has a bias your benchmark can't see.

A compression method that costs four tenths of a point on a benchmark average looks free. The average is the trap: it blends the capabilities compression leaves intact with the ones it quietly destroys — and the ones it destroys are instruction-following and long-context retrieval, the capabilities you shipped the model to have.

→
2026.10.23 OPERATIONS

A bad GPU can corrupt your training run for days without ever raising an error.

The mental model says a faulty accelerator throws an exception, health checks quarantine it, and a restart heals the job. Silent data corruption does none of that: it perturbs gradients with no error code, passes fleet screening, and your restart re-applies the bad update.

→
2026.10.23 TRAINING

The same bug breaks your eval numbers and your RL run.

An eval score that will not reproduce and an RL run that quietly collapses look like two unrelated problems — one a measurement nuisance, one a training-recipe failure. They are the same defect: floating-point non-associativity, surfaced by kernels that are not batch-invariant. The fix is a precision and kernel decision, not a hyperparameter.

→
2026.10.23 EVAL

'Lossless' speculative decoding is a property of the math, not your deployment.

Speculative decoding is provably distribution-preserving — a theorem about the algorithm. But you do not deploy the algorithm; you deploy a batched implementation of it on a GPU. Batched implementations have routinely violated the equivalence and emitted corrupted text, the speedup shrinks or inverts under load, and even a correct implementation is not bitwise-reproducible. Verify the equivalence; treat the speedup as measured.

→
2026.10.23 OPERATIONS

Your one-pool inference server is paying a prefill-decode interference tax it can't see on the throughput graph.

One batch posts a record tokens/sec while a single prefill chunk stalls every in-flight decode behind it. The number that pays the bill is goodput — SLO-compliant requests/sec — and it can be several-fold lower. The fix is a capacity-planning decision, not a kernel tweak.

→
NEW ENGAGEMENT · INTAKE

Tell us about it.

The more specific you are, the more useful our first reply.

SERVICE AREA
↩ ENCRYPTED IN TRANSIT
ASK THE FIELD NOTES BETA