mljourney

Reasoning models and standard language models are not competing products — they are different tools with different strengths, costs, and failure modes. Choosing between them is a matter of matching the tool to the task, not a one-time decision about which technology is superior. The practical question is: given a specific task, does it benefit … Read more

Reasoning models are language models that think before they answer. Instead of predicting the next token immediately from the input, they generate an extended internal monologue — working through a problem step by step, checking their own logic, backtracking when something does not hold, and only then producing a final response. This is what “chain … Read more

OpenAI’s o-series models replaced the traditional “scale up the base model” approach with an explicit reasoning step: the model thinks through a problem before producing an answer, spending more compute at inference time rather than training time. o3 and o4-mini are the current generation of this series. They differ not just in size but in … Read more

When DeepSeek released R1 in January 2025, it was the first open-weights reasoning model to match OpenAI o1’s benchmark performance on mathematics and coding — and it arrived at a fraction of the training cost, with weights anyone could download and run locally. The comparison between DeepSeek R1 and OpenAI o1 is significant not just … Read more

Reasoning models perform substantially better than standard language models on hard coding tasks. On SWE-bench Verified — a benchmark that evaluates models on real GitHub issues from major open-source repositories, where the model must write code changes that make tests pass — the performance gap is striking. Standard GPT-4o class models score around 33–38%; o3 … Read more

Budget forcing is a technique for constraining how much a reasoning model thinks before answering. It works by explicitly limiting the number of thinking tokens the model can use — either by setting a hard token cap, by injecting a prompt at a specific point in the thinking trace, or by using a model API’s … Read more

Test-time compute scaling is the observation that spending more compute during inference — having a model think longer rather than answer immediately — improves output quality in ways that parallel the improvements from spending more compute during training. It is the key insight behind reasoning models: instead of making the model larger or training it … Read more

DeepSeek released R1 alongside a family of distilled smaller models — ranging from 1.5B to 70B parameters — specifically so practitioners could run powerful reasoning locally without a data centre. The choice between 14B, 32B, and 70B is not simply “bigger is better”: it involves your hardware, your tolerance for latency, and how much of … Read more

Parameter-efficient fine-tuning (PEFT) methods adapt a pretrained large model to a new task by training only a small fraction of its parameters rather than the full model. The most widely used method is LoRA (Low-Rank Adaptation), which freezes the original weights and adds trainable low-rank matrices to specific layers. Hugging Face’s PEFT library provides a … Read more

FlashAttention-2 is a rewritten attention algorithm that computes the exact same result as standard attention but 2–4x faster and with memory usage that scales linearly with sequence length instead of quadratically. It achieves this by tiling the attention computation to stay within GPU SRAM rather than reading from the slower HBM memory repeatedly. For anyone … Read more