Artificial Intelligence

TL;DR Composer 2.5 is Cursor’s most capable in-house coding model yet, built on Moonshot’s open-source Kimi K2.5 checkpoint with about 85% of total training compute spent on Cursor’s own continued pretraining and RL The model is purpose-built for the agent loop inside Cursor - long-horizon tasks, hundreds of tool calls, multi-step instructions - rather than as a general-purpose chat model Cursor claims parity with Claude Opus 4.7 and GPT-5.5 on its own CursorBench v3.1 (63.2%) and a strong 79.8% on SWE-Bench Multilingual Pricing is dramatically lower: $0.50 / $2.50 per million input/output tokens on the default variant, with included usage doubled for the first week Together with SpaceXAI, Cursor is now training a much larger successor model from scratch on Colossus 2 with around 10x the compute - so 2.5 is a waypoint, not the endgame For a while, Cursor was an IDE wrapped around someone else’s models - Claude, GPT, Gemini. That story has shifted. With Composer 2.5, released this week, Cursor has shipped its most capable first-party coding model yet, and it is a serious enough piece of work that it deserves real consideration as a daily driver rather than a budget fallback. ...

A common dismissal of modern AI goes like this: “It is just a fancy autocomplete. It memorises text and stitches it back together. There is no real understanding, only retrieval.” It is a comforting story, and it has the shape of a critique that ought to be true. But spend enough time with frontier systems and a different picture starts to form. The thing that large models actually seem to be good at is not memorisation. It is something stranger and arguably more important: the formation of analogies, the combination of distant concepts, and the generation of conceptual relationships that were not explicitly present in any one place in the training data. ...

Dario Amodei is one of the few frontier-lab CEOs whose public talking points have not changed materially in five years. The same message he gave to small audiences in 2021 - that powerful AI is coming faster than people think, that the safety problem is real, and that the companies building it have an obligation to do so carefully - is the message he is giving to Congress and Davos in 2026. The thing that has changed is that he now runs the company most aggressively turning that message into a commercial position. ...

AlphaFold’s release in 2021 was the AI-for-science moment that broke through to the general public. A computational solution to a 50-year-old problem in biology - predicting protein structure from sequence - that produced a tool used by hundreds of thousands of researchers. The narrative around AI-for-science crystallised: deep learning would produce a series of similar breakthroughs across scientific domains. The 2026 reality is more interesting and less clean. AlphaFold-class breakthroughs have been rarer than the early narrative suggested. But AI has spread across scientific practice in subtler ways that, in aggregate, have done more to change how science is actually done than the few headline breakthroughs. ...

For most of the AI conversation in 2024 and 2025, the binding constraints on the build-out were chips and capital. By 2026 the conversation has shifted, and the constraint that gets discussed most seriously inside the hyperscalers is electricity. Not the cost of electricity. The actual physical availability of electrons - at gigawatt scale, in the places where the data centres need to be, on the schedule the model labs need them to be. The story does not have a single villain or a single number, but it has a shape, and the shape is becoming the story of the second half of the decade. ...

For most of the last decade, talking about AI hardware meant talking about Nvidia. In 2026 that has stopped being true at the inference layer. Three companies - Cerebras, Groq, and SambaNova - have built genuinely different chips around the same insight: that the workload economics of running models in production are not the same as the workload economics of training them, and that the chip architecture should follow the workload. The bet has been right enough that Nvidia has now licensed pieces of it. ...

For most of the LLM era the open-weight story was framed as a trailing one. Open models were cheaper, smaller, and a generation behind. That framing has not survived 2026. The gap between the best open-weight model and the best closed model is now narrow enough on most workloads that the choice is no longer “settle for less” - it is “decide what you actually need.” TL;DR Open weights have closed the headline gap. Top open-weight models are within striking distance of closed frontier models on reasoning, coding, and general knowledge benchmarks. The economics changed first. DeepSeek’s R1 made it credible that a frontier model could be trained for tens of millions, not billions - and that the weights could be released for free. Llama, Mistral, Qwen, and DeepSeek lead on different axes: Llama for broad ecosystem support, Mistral for European deployment and tool use, Qwen for multilingual and long-context work, DeepSeek for raw reasoning. Inference flexibility is the underrated win. Open weights mean you can run on your own hardware, fine-tune freely, and avoid surprises from a closed provider’s roadmap. The remaining closed-model advantages are real but narrowing - agentic depth, multimodal performance, and the polished tool-use stacks around them. Where the gap actually is in 2026 Benchmarks are imperfect, but the picture they sketch is consistent. On standard reasoning suites - MMLU, GPQA, MATH - open-weight models are within a few percentage points of the closed frontier. On coding - HumanEval, SWE-Bench - the gap is similar. On long-context retrieval, the gap is mostly gone. ...

TL;DR Document understanding is the unglamorous killer application - invoices, contracts, and scanned PDFs that were painful to extract data from are now tractable without dedicated pipelines Vision models still under-deliver on precise spatial reasoning, object counting, and subtle medical or scientific imagery - these remain jobs for specialist models Audio is the modality with the most upside: beyond transcription, it carries tone, pace, and hesitation that text loses, enabling fault detection, emotional analysis, and richer inputs The teams getting real value treat multimodal as an invisible enabling capability within a workflow, not a feature to demo - and they verify high-stakes outputs just as they would text The right question when evaluating multimodal is not “can we use this” but “what specific user problem becomes tractable that previously was not” When the first multimodal frontier models shipped, the demos were genuinely impressive. A photo of a fridge interior with the model suggesting a recipe. A handwritten napkin sketch becoming working code. A short audio clip of a meeting being transcribed, summarised, and structured. It looked, briefly, like the boundary between modalities had collapsed and we were entering a new regime in which models could reason fluidly across text, images, and sound. ...

TL;DR Prompt caching saves the computed representation of a prompt’s static prefix so subsequent requests reuse it rather than recompute it - cached tokens cost roughly 10% of normal input token prices The savings are highest when prompts have a long, identical prefix across requests - system prompts, tool definitions, and few-shot examples can make up 80-90% of total input cost The most common mistake is interpolating variables into the system prompt, which breaks caching silently; fix it by moving all static content to the top and dynamic content to the end Cache lifetimes are bounded (minutes to a few hours per provider) and any change to the prefix - including whitespace - creates a new cache miss Track your cache hit rate explicitly on every LLM dashboard; a dropping hit rate usually signals unintended prompt construction changes, and fixing it is the highest-leverage cost optimisation available If you build LLM applications for any length of time, you eventually notice that you are paying to have the model read the same instructions over and over again. The system prompt, the tool definitions, the few-shot examples, the structured output schema - all of it goes back into the model on every single request, and you pay for the input tokens every single time. For a chatbot doing one or two thousand requests a day this is annoying. For an agent doing tens of thousands of requests with long contexts, it is the dominant cost line. ...

Two years ago the way to make a model better was to train a bigger one. By the start of 2026 that recipe has stopped being the most interesting answer. The frontier has moved to a different lever - letting the model think for longer at inference time, generating intermediate reasoning, and only then producing the final answer. The category has a name now (reasoning models) and a family of products built around it. The interesting questions are no longer whether the trick works, because it clearly does, but when to reach for one, where it lands in production, and what the costs actually look like once the demo glow wears off. ...