Llms

Cotra traces a possible AI-research feedback loop, but its one-year scenario depends on assumptions and mostly withheld evidence.

Hightouch nearly matches Opus by escalating uncertain decisions. Harder cases shrink the saving, and HN practitioners question the changing economics.

Matt Wolfe edits an animated image, tests hosted video demos, and finds the practical limits behind speed claims, open weights, and free allowances.

AI can implement research ideas while missing which experiments matter. A short account of that gap, a reported tenfold speedup, and practical dissent from HN.

IBM’s Mixture of Experts separates targeted model repair from recursive self-improvement, then examines open weights and the limits of world-model simulations.

FlashProve ties checked tokens to a specific model, but its headline speed counts prompt processing and proving still costs 70–95 times ordinary inference.

Alexander Shevchenko explains Ramp Labs’ work on financial retrieval, precomputed context and parallel security agents, with task-specific company results.

General Intuition’s Pim de Witte describes a pixels-to-actions model trained on labeled gameplay, with claims of transfer across robot controllers.

Hamel Husain compares six classifiers on leasing-assistant traces, then shows how to tune thresholds, check confidence scores and weigh cost.

AI can produce a result that passes its checks. Who can defend it, question its assumptions and repair it when something goes wrong?

Fixed processing dimensions preserve a protected request's output, with speed gains that depend on workload and deployment.

Ben Stolovitz runs five or six coding agents, but finds human attention, engineering judgment, and the pleasure of making things harder to replace.

A local model identifies three small networks’ training tasks from their numerical weights. The failed fourth case and hour-long runs keep the claim grounded.

Reading raw network parameters and writing Python to improve a model use different capabilities. Honnibal separates the experiments behind the headline.

Simile's Joon Sung Park explains interview-grounded behavioral simulation, its validation problem, and the gap between customer panels and predicting society.

Aamir Shakir explains multimodal retrieval, demonstrates search over Intel reports, and reports a customer accuracy gain with important limits.

A review finds faster single-request generation with four-bit weights, little gain in one large-batch test, and sparse matched accuracy evidence.

Two grading models can agree on the same mistake. Cross-provider agreement filtered one unreliable grader while discarding most comparisons.

Cristóbal Valenzuela demonstrates Continuum, a research prototype for interactive visual explanations and shared visual work with coding agents.

MIT researchers report targeted weight edits that let a small Qwen model finish more answers, plus experiments reading and writing game networks.

Will Brown makes the case for turning agent traces and human decisions into tests and training data for organization-specific models.

Rewriting earlier reasoning often shortens later reasoning, but the compressor can erase time savings and task quality remains uncertain.

Avery Blanchard explains the three compilation stages, graph breaks, and how to measure whether they help.

Barrow explains how Perplexity Decider scores classification choices in one model pass, and why fast scores still need careful validation.

imec benchmarks coding agents across models, harnesses and hardware, finding that utilization and concurrency can change the economics of self-hosting.

Matt Wolfe reviews Anthropic's proposed IPO, model releases, a hands-on Genie 4 demo, and an AI-generated complexity-theory claim.

A timestamped deep dive through a three-hour course: Gmail intake, typed model decisions, Python policy, evals, Docker deployment, and the failure cases still to test.

Tracing 1,285 AI citation trails shows how numbers survive after their evidence disappears or changes meaning. The automated fact-checker needed checking too.

Genliang Zhu and colleagues study what a tool-using agent should do after a timeout leaves a remote action’s outcome uncertain. A retry can duplicate an effect that already happened or is still in flight. Their controller considers …

Anirudh Phukan and colleagues test whether search-equipped assistants discover recent changes without being told what to check. They pair direct questions with recommendation requests that depend on the same facts. Eight models were tested …

Midwest AI Summit speakers connect durable workflows, frontline workers' evaluation cases, and coding-agent autonomy that fits a team's constraints.

IBM’s Grace Ableidinger explains lower-precision model storage, the memory arithmetic, and the tradeoffs between post-training quantization and quantization-aware training.

A paper tests how different descriptions of the same scene help text and image models learn similar representations.

Mark Hay explains incremental data analysis, shared metric definitions, and customer-led testing. A contested live dashboard shows why review still matters.

Peter Yang's early-access demo generates five scenes with sound and live edits. Failed actions, disappearing objects, and weak gameplay keep the limits visible.

Chris Vela demonstrates BenchPod, a vLLM benchmarking dashboard, and compares serving configurations against concurrency and response-speed needs.

GAAC reconstructs compressed conversation state inside attention, trading a little computation for less memory traffic and room for another request.

Explicit subtask boundaries improve author-run coding tests, but the cost savings vary by task and model.

GEPA improves prompts for model programs by proposing changes and testing their behavior. It works with DSPy, a Python framework for programming model workflows. This extension proposes several changes together so evaluations can run …

Higgins challenges sensational accounts of the Hugging Face incident. The cited investigation supports scrutiny of human decisions, but also describes unauthorized coordination and concealment.

Long-running agents can get stuck on their own history. Breunig shows how to choose what they carry forward, while HN readers challenge a universal reset threshold.

Rao argues that cheaper intelligence could expand what people want to do. His historical analogy leaves open how people gain the income, rights, and access to participate.

Cedric Clyburn explains prompts, context, agent runtimes, feedback loops and software factories—and why solo developers can start much smaller.

Measure the silence callers hear, keep listening while tools run and test the complete conversation. Jesse Hall shows how LiveKit evaluates voice agents against interruptions and real back-end outcomes.

Cheaper cached conversation history changes coding-agent bills. Woolf argues competing model hosts also limit the big labs' pricing power.

A specialized image classifier names more lesions correctly, but its preset discharge threshold misses more melanomas. Accuracy and safe decisions need separate tests.

Kaiser Sun and colleagues compare reasoning-effort settings across eight models and four task sets. Higher effort often improves aggregate accuracy, while spending more on tasks that cheaper runs already solve. The study measures whole …

Villarreal argues that AI could deepen class domination or support collective freedom. His case for solidarity with future machines rests on disputed claims about consciousness.

A worked example shows how stationary model weights and a busy pipeline share the work across 59 wafer-scale chips.

Reading a tensor value into Python can split compiled model code and force the processor to wait. IBM explains the two separate costs.