Catastrophic Forgetting in LLM Fine-Tuning
Catastrophic forgetting in LLM fine-tuning: why sequential training erases skills, the published fixes from replay to OSFT, and the eval gate that catches it.
Read the articleNotes from
the work.
Guides, research notes and what we learn building models.
Catastrophic Forgetting in LLM Fine-Tuning
Catastrophic forgetting in LLM fine-tuning: why sequential training erases skills, the published fixes from replay to OSFT, and the eval gate that catches it.
Read the article
Continual Learning for AI Agents: The Model Layer
Continual learning for AI agents at the model layer: the agent's own traffic trains each new model version, scored on your evaluations before it goes live.
Read the articleContinued Pretraining vs Fine-Tuning vs RAG
Continued pretraining vs fine-tuning: what continuous pretraining is, how domain knowledge enters a model, and when retrieval fits better than training.
Read the article
Domain-Specific LLM Evaluation: A Worked Method
Domain-specific LLM evaluation step by step: a held-out set, a frozen protocol, thresholds set in advance and confidence intervals, from our QMSum study.
Read the article
What Is a Domain-Specific LLM? With Examples
A domain-specific LLM knows one field and serves many tasks in it. Sourced examples, how it differs from a task-specific model, and how the two combine.
Read the article
Foundation Models vs Task-Specific Models
Foundation models vs task-specific models: what each is, when a model trained for one job wins, examples you can use today, and how specialists combine.
Read the articleOpen-Weight vs Frontier Models: Cost and Quality
Open-weight vs frontier models: how far open models trail, why they cost less per token, and when a model trained on your task beats a rented frontier model.
Read the articleWhat Is an Applied AI Lab? Three Kinds of Lab
An applied AI lab trains models for work that already exists and earns from its first deployment. How it compares with frontier labs and neolabs, with examples.
Read the articleAI Research Labs in Australia: A 2026 Map
A sourced map of AI research labs in Australia as of October 2026: university institutes, CSIRO and commercial labs, each with one checked line.
Read the articleAI Startup Gross Margins and Inference Costs
AI startup gross margins averaged 45% in 2025 because inference costs scale with usage. Where the cost per task goes, and what a trained model changes.
Read the articleIs Fine-Tuning Dead? What the Labs' Exit Means
Has fine-tuning been debunked? OpenAI and Mistral are winding it down. Where trained open models still win, where prompting is enough, and our measured results.
Read the articleOpenAI Fine-Tuning Alternatives: Migration Guide
OpenAI fine-tuning deprecated: new jobs end 6 January 2027. What happens to your models, the alternatives, and how to migrate to trained open models.
Read the articleWhen Should Enterprises Build Custom AI Models?
A practical guide to deciding whether a repeatable enterprise workflow merits a custom AI model, using market data and measurable evaluation criteria.
Read the articleAgentic AI Cost Optimization: Cut Tokens per Run
Agentic AI cost optimization for engineers: where agent tokens go, how to trim tool output, cache the prefix, cap each run and track cost per successful run.
Read the articleCost per Task: Measuring AI Unit Economics
Cost per task turns an AI token bill into unit economics. How to calculate it, what an AI agent costs per month, and how to set a target and a ceiling.
Read the articleAI Token Budgets: How to Set One for Your Teams
How to set an AI token budget for your teams: allowances by role, showback, default models, coding agent limits and a cost-per-task budget for workloads.
Read the articleWhen a Custom Model Beats a Frontier API on Cost
Break-even math for replacing a frontier API with a custom model: the formula, a worked example at October 2026 prices, hidden costs and a safe migration path.
Read the articleHow to Reduce AI Token Costs in Your Company
A guide for leaders: find where your AI tokens go, take the quick savings first, then give your highest-volume task a custom model of its own.
Read the articleLLM Cost Optimization: 12 Techniques Ranked
Twelve LLM cost optimization techniques ranked by savings, effort and quality risk, with October 2026 list prices for caching, batching and model tiers.
Read the articleLLM Routing Strategies to Cut Frontier API Costs
Six LLM routing strategies compared on cost, latency and risk, plus the shadow, canary and ramp playbook for moving traffic to a cheaper model safely.
Read the articleWhy Are AI Costs So High? What Leaders Can Do
AI token prices keep falling, yet AI bills keep rising. Why costs climb, how to tell if the spend pays off, and what leaders can do, with October 2026 data.
Read the articleAMD acquires Taalas: model-specific AI silicon
AMD agreed to buy Taalas on 6 August 2026, eight months after Nvidia licensed Groq for a reported $20B. Both deals say inference is splitting from training.
Read the articleLFM2.5 sizes compared: 230M vs 350M vs 1.2B on-device
Liquid AI ships LFM2.5 at three sizes. Benchmarks, export sizes, and what we learned fine-tuning two of them into shipped products, including the failure no metric caught.
Read the articleFine-Tuning for Indie Devs: Inside the $10 Ertas Lite Plan
The cheapest way to train a custom AI model you own. What $10/mo unlocks for hobbyists and indie devs, which models you can train, and where it runs out.
Read the article