Llms

Kaggle's benchmark audits expose fragile scores. LinqAlpha's MandateBench tests the reasoning behind a full investment memo, with early results from ten tasks.

GEPA evolves prompts for model programs, including workflows built with DSPy. This study tests its merge operation, which combines candidates descended from a common starting instruction. The comparison spans five datasets, two models and …

Harneet Kaur tests two models on iHeartMedia. Checking each reasoning layer makes their forecasts converge, but agreement leaves accuracy unproven.

Aman Singh explains how Arca coordinates days-long client work with event signals, persistent memory, shared plans and human approvals.

Matt Wolfe’s weekly roundup spans Claude Duet, software-security incidents, editable vectors, agent access and cautiously framed medical-AI results.

Jump Trading’s compute leads explain bespoke GPU contracts, startup prepayments and why a price hedge cannot deliver a cluster.

Separate model sessions recover messages from rendered drawings, while the decoding tools leave a visible trail.

Joseph Kim describes licensed financial data, firm templates and inspectable citations, then separates today’s product from his 2027 predictions.

Carmen Li explains reserved GPU procurement, provider verification and forward pricing, separating delivered capacity from financial hedges.

Freddy Lecue describes Wells Fargo’s transaction-trained transformer, its data-tokenization work and reported fraud gains over the bank’s XGBoost baseline.

Anthropic and Theorem report checked proofs for 239 test entry points across 123 libraries. The public release, machine model and patching process need separate scrutiny.

ModelCompare's small experiment finds a familiar refrain about tribal loyalty and rationalization. Its broad answer categories and leading prompt limit what that agreement means.

San Francisco Compute’s Evan Conrad argues that deposits, delivery delays and uncertain demand form one financing problem. Tradable capacity needs reliable cluster operations behind it.

MacPaw researchers test four families of model-serving optimizations on Apple Silicon. Their measurements show why a server-GPU speedup needs checking against the model, software and request pattern you actually use. Pruning the KV cache, a …

A small model learns terminal behavior, but removing one layer preserves plausible output while destroying its memory.

Grace Ableidinger follows a model request through tokenization, routing, batching, prefill and decode, and explains what first-token and inter-token latency measure.

Abhishek Barali and Aayush Paudel test what happens when an extraction schema requires more answers than a document contains. Models often fill the gap with genuine text of the wrong kind. Across eight downloadable models and 558 documents, …

Text-only editing damaged some correct transcripts. Adding audio reduced that damage for Gemini, but also reduced repairs to its own drafts.

Directable voices, persistent AI worlds and 3D reconstruction, with the hardware limits and demo-only claims kept in view.

MiniMax's Morgan Suo connects faster model internals to the cost, accuracy and delivery time of a research report an analyst can actually use.

Thomas Hikaru Clark, Matias B. Martin and Kyle Mahowald examine why AI prose can feel difficult despite being grammatically valid. Their study finds features associated with reading difficulty across model-generated conversations, coding …

Oliver Kennon explains how serving software, reliability and security change the cost of useful AI output, even on the same GPU.

A study of ChatGPT, Claude and Gemini shows how saved malicious instructions can affect later chats. Memory injection means storing an instruction that returns as context in future sessions. The authors tested ChatGPT Instant 5.3, Claude …

Invented statistics acquired real citations. A source link can preserve a falsehood as easily as it can verify a fact.

Single-block profiling finds that attention software leaves newer workstation GPUs underused, while proposed fixes remain a systems research agenda.

Moshi founder Anekha Sokhal argues that AI should gather and check market information while investors keep the decision. She describes a six-agent system, traceable claims and the hidden cost of losing apprenticeship.

AI can turn shared knowledge into private wealth while overwhelming the reviewers and maintainers who keep it alive.

Benj Edwards makes songs and games for an audience of one. Their private value can be real while the consequences for creative workers remain unresolved.

Request-level logs make coding agents easier to debug. The HN discussion asks whether that requires putting another service in every request's path.

A two-model GEPA and SkillOpt study separates answer correctness, formatting, server failures, waiting time and token use.

IBM Technology's Kaoutar El Maghraoui explains the memory wall, how to locate a workload on a roofline chart, and why batching helps language-model decoding.

A visual explanation of how language models remember, built around the interference that makes recall fail.

Ben Carsley explains context-specific writing styles, the limits of prompt-only personalization, and shared model serving with a small adapter per person.

Mercor maps AI tests against American jobs and finds large blind spots. The coverage gap measures how little we know about models at work.

Wolfe’s weekly roundup contrasts Krea’s editing workflow, downloadable models, one-shot coding demos and the smaller updates behind product launch headlines.

Generating research is not the same as advancing understanding. A physicist asks what AI is accelerating—and who gets left checking the answers.

What if the apparent tangle inside language models comes partly from bad microscopes? A case for studying the original neurons, with important limits.

Hamel Husain argues for validating AI judges against human labels before adding another model provider: measure precision and recall, then optimize cost.

A demand for public evidence meets a disagreement about whether better AI can quickly become cheaper goods and automated jobs.

Martin Keen maps four context failures to four fixes: validate mistaken notes, compact long histories, curate tools and remove superseded information.

An eight-GPU serving study fixes a cache bottleneck, while retaining the faster median of its recomputation baseline.

A trained memory interface feeds compact numerical representations to a frozen model. Its agent experiment still retains the full history.

A second model runs checks after completion, then sends failures back while preserving the working agent's state.

CODA tests jointly trained models and skill libraries after their instructions change. Even paraphrases can break previously successful tasks.

Slow reasoning improves a fast game-playing agent when it happens between attempts, with validation deciding which instructions survive.

A single-GPU study preserves returning sessions while loading the next expert, but its extreme capacity claims need clarification.

HybridAgent mixes internal vector-based computation with text for tools and inspection. Matched-model tests show stronger search results, but the cost savings cover estimated generation compute.

Peter Yang compares code-built scenes, games, storytelling, websites, and writing—and finds different winners for different tasks. These are selected demos, not a controlled benchmark.

A controlled study separates finding a skill, opening its instructions and completing the task. Choosing skills before execution helps, with costs.

Copies of one old claim can crowd out a correction. A memory study tests source tracking, and shows when filtering makes answers worse.