AI Research
38 essays tagged with AI Research.
AuthorizationBench: Success Does Not Grant Authority
A deep assessment of AuthorizationBench v2: benign pressure, consultation timing, statistical limits, a pinned release audit, and how to evaluate authorized agent work.
Nova 2.5 Sonic and Strands Bidi: Test the Conversation Across Reconnection
A research assessment of Amazon Nova 2.5 Sonic and Strands Bidi GA: speech reasoning, renewed connections, asynchronous tools, and the economics of a resolved call.
Haiku 5.5: Small Models Make Delegation a Partitioning Problem
An assessment of Haiku 5.5's adaptive effort and pricing breakpoint, with experiments for deciding which work belongs in a small-model subagent.
Cohere's Coding-Agent Study: Buy Information Before Buying More Implementers
Cohere's October 14 agent-team study motivates a sharper allocation rule: measure what workers teach the lead, separate partial reward from solved tasks, and include integration cost.
DeepSeek Search and Flash-1017: The Price of Finding the Right Evidence
DeepSeek's October 17 search APIs and Flash update create a testable question: do cheaper research workflows preserve evidence quality and verification?
Argon in Production Code: What Shipped After the Overnight Run
Google's October 13 Argon rollout and ADK case study reveal how to evaluate autonomous performance work through accepted changes and held-out workloads.
Muse Spark Open: Active Parameters Do Not Size the Server
Meta's October 15 open-weight release creates a deployment choice. Assess the full weights, KV cache, quantization, parser behavior, and utilization before comparing costs.
Muse Wonder: Test Memory and Persistence Separately
Meta's Muse Wonder release pairs Active Memory with Active Focus. A research assessment of their interaction, benchmark methodology, and a practical ablation study.
Microsoft MAI-Base-2 and Code Flash: Three Different Meanings of Efficiency
An assessment of MAI-Base-2 pretraining and MAI-Code-1.1-Flash: scaling experiments, useful-code economics, local inference, and the evidence each decision needs.
MiniMax M3.1 Flash: Cheap Tokens Change the Experiment, Not the Acceptance Test
A primary-source assessment of MiniMax M3.1 Flash: inference economics, reasoning controls, API compatibility gaps, and a concrete migration experiment.
Mistral Large 4 and Timestral: Two Models, Two Different Scaling Decisions
Mistral's October releases separate general agent reasoning from numerical forecasting. How to evaluate Timestral, preserve issue-time covariates, and price a hybrid workflow.
Kimi's Cache and Search APIs: Buying Reuse Without Losing Evidence
A research assessment of Moonshot's September cache and search releases: write pricing, TTL break-even, evidence selection, and long-session evaluation.
Nemotron 3.5 Super VL: The Specialization Bet Behind the Speed Claims
A source-backed assessment of NVIDIA's October 15 Nemotron 3.5 Super VL and Bolt releases: multimodal specialization, cache reuse, quantization, reproducibility, and accepted-task economics.
OpenAI's October APIs: Faster Decisions Need a Latency Budget
An assessment of native Decisions in GPT-Live and Sol Ultrafast: separate network, generation, and commitment latency before choosing an architecture.
OpenAI's October Science Releases: Verification Capacity Is Infrastructure
The October mathematics collection and Rosalind Academic Plan expand candidate research and access. Assess the human understanding, validation, and stewardship needed afterward.
Qwen's E-Commerce Bench: A Profitable Agent Can Still Be a Bad Merchant
A critical reading of Qwen's long-horizon commerce benchmark: deterministic economic decisions, residual variation, fraud, cash timing, and a proposed memory ablation.
Grok Voice 3.0: Fast Speech Is Only One Part of a Resolved Call
An assessment of Grok Voice 3.0's October release: benchmark denominators, audio playback versus tool execution, interruption tests, and cost per durable resolution.
GLM-Legal: Domain Specialization Meets the All-Pass Test
A research assessment of Z.ai's GLM-Legal: model–harness co-training, independent benchmark evidence, all-pass versus criteria scores, and specialist evaluation.
Bedrock Managed Agents: What Moving the Harness Into AWS Actually Changes
A research assessment of AWS's OpenAI-powered managed-agent preview: harness placement, microVMs, execution roles, session lifetime, and deployment acceptance.
Claude 5.5: Efficiency Gains Need a System-Level Evaluation
A research assessment of Opus 5.5 and Sonnet 5.5: workload economics, safeguard fallbacks, benchmark comparability, and a migration experiment.
Cohere Embed 5: Retrieval Quality Is Only the Start of an Evidence Pipeline
A research assessment of Embed 5 and North Small Translate: shared embedding spaces, index migration, multilingual evidence, deployment limits, and retrieval evaluation.
DeepSeek V4.1 Flash: Cache Economics Meet Model Migration Risk
A research assessment of DeepSeek V4.1 Flash's asymmetric architecture, cache claims, native vision, and conflicting V4 Pro migration notices.
Gemini 4 Argon and Private Memory: Two Different Trust Questions
A research assessment of Gemini 4 Argon's restricted rollout and Google's private server-side memory design, separating capability, confidentiality, and memory correctness.
Muse for Small Business: The Approval Is Part of the Product
A deeper assessment of Meta's September 29 Muse expansion: connector scope, meaningful approvals, cross-app recovery, and the economics of owner attention.
Microsoft's New MAI Voice Stack: A Partial Transcript Is Not a Final Instruction
A research assessment of MAI-Transcribe-2-Streaming and MAI-Voice-2.1: latency, revisable speech, speculative tool use, and a voice-agent evaluation plan.
Mistral Vibe: Editable Memory and the Difference Between an App and a Snapshot
A research assessment of Mistral's September Vibe updates: Agentic Knowledge Base, spreadsheets, Mini App Canvas, migration, and artifact verification.
NVIDIA OpenShell and Sentry: Containment Needs an Independent Boundary
A research assessment of NVIDIA's September 28 Open Agent Safety Platform: software enforcement, an out-of-band watchdog, assurance boundaries, and falsifiable containment tests.
OpenAI After DevDay: Cheaper Intelligence, Persistent Agents, and the Cost of Accepted Work
A research assessment of GPT-6.1 Sol, Dots, and the DevDay platform changes: what the evidence supports, where costs move, and how to evaluate persistent work.
Qwen-Planner-Agent: What Model–Harness Co-Evolution Can Actually Prove
A critical assessment of Alibaba's September 24 Qwen-Planner-Agent preprint: hybrid environments, development-set feedback, attribution, and a stronger transfer experiment.
Grok 4.7 and Team Bots: Shared Expertise Needs Explicit Authority
A research assessment of Grok 4.7 and Team Bots: model–harness specialization, shared credentials, private memory, and evidence for team productivity.
What the Evidence Actually Shows About AI Automating AI Research
A source-backed examination of AI research automation: company disclosures, METR evaluations, developer productivity, PostTrainBench, and the gaps between capability and acceleration.
The Bottlenecks Between Recursive Improvement and an Intelligence Explosion
The compute, data, research-return, and verification constraints between AI self-improvement and an intelligence explosion, with a checked sensitivity analysis of the CASP model.
Who Benefits When the Machines Accelerate Discovery?
Who gains from faster AI discovery? A research-backed analysis of economic adoption, concentrated power, institutional capacity, and preparation across three scenarios.
Can Oversight Keep Up With AI Research?
How to evaluate oversight as AI research accelerates: intervention timing, independent evidence, agent reliability, and controlled promotion of research outputs.
When AI Improves the Process of Improvement
How automated AI research could create a recursive feedback loop, what existing experiments demonstrate, and how to test whether improvements truly compound.
Welcome, Astra: What GPT-6 Changes About Intelligence at Work
A research analysis of OpenAI's GPT-6 Astra: benchmark evidence, scientific discovery, tool use, safety, economics, and an evaluation plan for real work.
Anthropic Research Through August 4: Model Minds, Scientific Agents, Robotics, and the Economy
A source-backed review of Anthropic's newest work in interpretability, alignment, cryptanalysis, biology, robotics, autonomous systems, values, and economic measurement.
OpenAI’s August 2026 Science Work: From Benchmarks to Verifiable Discovery
What OpenAI's August mathematics, scientific-computing, biology, and evaluation work says about AI as a research collaborator.