Cover image

Forty-Eight Features Gone, Fifty-Nine Too Far

TL;DR for operators Removing features from a production recommendation model is a release decision, not simply a feature-ranking exercise. Pinterest’s FORTE framework1 makes feature retirement repeatable through daily importance monitoring, joint feature testing, retraining, online validation, and dependency tracking. The evidence explains why these stages matter. In Search Relevance, removing 59 features passed offline evaluation but failed a secondary online guardrail; removing 48 succeeded. In a separate interaction experiment, individually favorable feature changes summed to a +6.554% Closeup AUC effect, but their jointly measured effect was only +3.084%. ...

November 17, 2026 · 8 min · Zelina
Cover image

Five Model Widths, One Shared Trunk

TL;DR for operators A model platform offering five inference-capacity tiers ordinarily faces a choice between separately training five models and sharing parameters while varying the width trained at each step. ShellElastic1 introduces a third arrangement: all five supported widths receive gradients through one shared full-width transformer trunk computation. Its architectural guarantee is exact, provided the supported widths and operations satisfy specific constraints. Its companion optimizer parameterization, mu-Elastic, addresses the different update scales encountered by nested widths. Across three tested model scales, the combined method improves mean validation loss over the evaluated elastic alternatives. Independently trained models still achieve better average loss, however, at five times the aggregate optimizer steps. For training-platform operators, the central question is whether the shared architecture delivers acceptable quality across all service tiers at lower total GPU cost, not merely fewer training steps. ...

November 6, 2026 · 9 min · Zelina
Cover image

When the Federated Fix Makes Transcription Worse

TL;DR for operators A federated training algorithm designed to control differences between users can improve one multilingual speech model while degrading another. In a 316-client benchmark, FedProx reduced Whisper+EuroLLM’s word error rate (WER) from 13.30% to 12.17%, but increased Whisper+TinyLlama’s from 14.15% to 14.99% under corresponding frozen-encoder conditions. Model components mattered just as much. A multilingual language decoder performed better than its alternative, while enabling acoustic-encoder training without appropriate learning-rate tuning made recognition substantially worse. The best reported federated result, 11.70% WER, came from FedAvg with a carefully tuned trainable encoder—not from FedProx. ...

October 22, 2026 · 8 min · Zelina
Cover image

Commit First, Fail First: Qualify the Judge Before It Steers Optimization

TL;DR for operators If an LLM judge controls which prompts, agents, models, or generated outputs survive repeated optimisation, test the judge independently before trusting its scores. In Idil Gozel’s experiments, ordinary best-of-N search found many Python solutions that scored highly under a shipped LLM-judge configuration while failing hidden correctness checks.1 Requiring the judge to solve the problem before seeing the candidate fixed that failure completely on one task: divergence fell from 90/96 and 93/96 candidates across two seeds to 0/96 in both. ...

September 22, 2026 · 7 min · Zelina
Cover image

No Cluster Is an Island: ScaleAcross Explorer and the Geography Tax of AI Training

GPUs used to have a simple business story: buy more, wire them well, train bigger models. That story is not false. It is just starting to resemble a children’s book. The adult version has buildings, regions, power constraints, optical links, oversubscribed networks, packet loss, pipeline bubbles, model chunks, microbatches, and a quiet question with a very expensive answer: when the GPUs no longer fit comfortably inside one data center building, how should the training job be split? ...

June 5, 2026 · 18 min · Zelina
Cover image

Rank and File: BoostLoRA’s Case for Smarter Fine-Tuning

Opening — Why this matters now Enterprise AI is entering its less glamorous phase: not the demo, not the keynote, not the charming chatbot that answers three curated questions correctly, but the operational grind of making models behave reliably inside messy workflows. That grind usually runs into a familiar triangle. Full fine-tuning is powerful but expensive, operationally heavy, and often risky when the training set is narrow. Parameter-efficient fine-tuning, especially LoRA-style adaptation, is cheaper and easier to deploy, but the smallest adapters can hit a ceiling. Meanwhile, the business user does not care whether the adapter was elegant. They care whether the model stops making the same costly mistakes in invoicing, compliance review, customer support, code generation, or scientific triage. ...

May 4, 2026 · 13 min · Zelina
Cover image

When Right Meets Wrong: Teaching LLMs by Letting Their Mistakes Talk

Training a reasoning model is often treated like running a classroom with a very impatient teacher: give the model a problem, let it produce several answers, mark each answer right or wrong, and push the policy toward the winners. That is already useful. It is also slightly wasteful. Because in a real classroom, the wrong answers are not just trash to be swept off the floor. They reveal what the student misunderstood. They show which shortcuts are tempting, which algebra step keeps breaking, and which false pattern looks suspiciously persuasive. A good teacher does not only praise the correct solution. A good teacher puts the correct and incorrect attempts side by side and asks: what exactly changed? ...

March 16, 2026 · 16 min · Zelina
Cover image

When Failure Pays Dividends: Recycling Reasoning in RLVR with SCOPE

Failure logs are usually where AI teams put the evidence that training was expensive. A reasoning model tries a problem. It gets most of the chain right. Then, near the end, it makes one bad algebraic turn, chooses the wrong case, or quietly invents a rule that mathematics did not approve. Under standard reinforcement learning from verifiable rewards, that rollout receives the same score as nonsense: zero. The model may have climbed nine floors and tripped on the final step; the reward system marks it as indistinguishable from someone who never entered the building. ...

March 2, 2026 · 15 min · Zelina
Cover image

When the Answer Matters More Than the Thinking

Answer. In most business systems, that is the part users actually care about. The approval decision. The risk label. The final invoice category. The recommended next action. The tidy little field that decides whether the workflow moves forward or someone opens a Slack thread titled “Why did the AI say this?” Yet much of modern LLM fine-tuning treats that answer as just another slice of text. Worse, when supervised examples include long chain-of-thought explanations, the final answer may become the shortest and least dominant part of the training objective. The model learns to produce a convincing trail of reasoning, but the tiny destination at the end receives comparatively little optimization pressure. Very elegant. Also slightly absurd. ...

December 26, 2025 · 15 min · Zelina