The latest developments in how AI is built.
ModelsAnthropic’s October release brings adjustable reasoning to its low-cost model tier. The business question is what work can be delegated without losing the evidence that matters.
ModelsThe new multimodal preview puts visual grounding under scrutiny: can a model identify the right evidence before reasoning about it?
ModelsThe new model makes inference efficiency a selling point. The important distinction is between estimated computation and the cost of completed work.
ModelsGoogle’s October 6 release broadens embeddings beyond text. The model-building opportunity is a shared representation; the challenge is deciding what similarity should mean.
ModelsV4.1 Flash makes architecture a business variable: the cost of remembering a long conversation may matter as much as the cost of answering it.
ModelsOpenAI’s September release raises a more useful purchasing question: how much frontier capability does a task actually need?
ModelsWhole-body control attracts attention. The deeper model-building question is how much learned behavior survives a change of robot.
ModelsA steerable VLA points toward a different model-building advantage: reusing learned skills through richer context, rather than collecting every task again.
DataCapture hardware and reconstruction models turn human activity into candidate training data. The hard part is preserving what transfers to a different body.
DataPegasus 1.6 brings first-person footage into a structured labeling pipeline. The test is whether it saves effort without teaching models the wrong actions.
DataFor models that take actions, a dataset increasingly needs states, tools and verifiable outcomes. The design of that environment determines what the model learns.
DataAn October launch makes a consequential claim: companies may need better-selected tokens before they need a bigger model.
DataNeMo Data Designer’s recent research and software updates move attention from generating examples to controlling what those examples teach.
TrainingReinforcement learning must keep challenging a model as it improves. The engineering problem is producing useful attempts, trustworthy rewards and a curriculum that does not stand still.
TrainingNemotron 3 Ultra shows why cheaper pretraining depends on knowing which calculations cannot afford to be cheap.
TrainingFaster chips cannot remove every bottleneck. The new attention kernels show why training software determines how much research a GPU budget buys.
TrainingA specialist-model result suggests that knowing what a correct answer means can matter more than adding another layer of prompts.
InferenceSeparating prompt processing from generation changes the capacity question: how much work can finish within a useful deadline?
InferenceThe latest cache-compression results expose two distinct gains: faster generation and room for more live sessions.
InferenceInco AI’s parallel drafter targets more accepted tokens per model pass, with gains that depend on the workload it serves.
InferenceBonsai 2 27B makes deployment location a choice to evaluate, while aggregate benchmark retention leaves important questions unanswered.
EvaluationsA contracting benchmark connects model training to business rules. Its partial-credit results expose the distance between an improving agent and a workflow a company can trust.
EvaluationsGDP.xlsx tests whether models can reconstruct a workbook’s logic. Its implications reach back into training data, tool use, and how enterprise capability is measured.
EvaluationsFrontier models beat a small human baseline on structured accounting tasks. The business question is what remains outside the benchmark.
EvaluationsArtificial Analysis adds a hallucination gate to its legal-agent benchmark. The change exposes a model-development problem: covering every requested point and grounding every claim are different skills.
EvaluationsA model that declares success is not necessarily finished. Legacy-Bench shifts attention from code generation to the evidence needed to trust a repair.
