Intel
Page 47
Designing AI-Intensive Applications - swyx
As applications get more AI-intensive, the ratio of user input to model calls moves from 1:1 in a chatbot to 1:100 in deep research or Codex.
Why it mattersswyx frames a mental model for building AI-intensive apps as they shift from chatbot-style 1:1 interaction to deep-research and ambient agent patterns (1:100, 0:n).
videoAI Engineer
From GPT-2 to gpt-oss: Analyzing the Architectural Advances
An in-depth technical article breaking down the architectural design choices in OpenAI's gpt-oss-120b and gpt-oss-20b open-weight models, comparing them against GPT-2 and Qwen3.
Why it mattersIf you're evaluating or building on OpenAI's gpt-oss models, this walks through their exact architectural choices (MoE, sliding-window attention, MXFP4 quantization) with clear comparisons to GPT-2 and Qwen3.
articleSebastian Raschka, PhD
The Future Isn't Model Agnostic
An argument against building for model portability.
Why it mattersMakes the case that abstraction layers for hot-swapping LLMs buy little, because prompts, evals and product feel are all model-specific — the leverage is in knowing one model's edge cases deeply.
articlefly.io
GPT-5
GPT-5 is OpenAI's flagship AI model that automatically selects the right reasoning level for each task and proactively takes action on vague prompts.
Why it mattersGPT-5's default toward autonomous, multi-step action on vague prompts changes how you scope tasks for it.
articleEthan Mollick
Titan The Model Factory S Furnace
Poolside's pretraining codebase, and the reasoning that produced it.
Why it mattersThe priorities that shape a training codebase invert between a few nodes and 10,000 GPUs. Debuggability and memory headroom give way to reliability and automated ablation, and Titan is the record of that transition.
articlePoolside editorial sitemap
Andrew Ng@AndrewYNg
A course with Anthropic, taught by Elie Schoppik, on using Claude Code past the basics: coordinating multiple subagents on different parts of a codebase, autonomous issue-to-merge GitHub workflows, notebook-to-dashboard cleanup, and Playwright over MCP for visual debugging.
Why it mattersCovers the Claude Code patterns most users never reach: parallel subagents across a codebase, agent-driven GitHub issue-to-PR workflows, and MCP-connected Playwright so the agent inspects and fixes its own UI output.
Mcp Ui Breaking The Text Wall
Shopify's open extension to the Model Context Protocol lets MCP servers return interactive UI — product selectors, galleries, cart flows.
Why it mattersMCP responses are text-only by default; MCP UI defines how a server can return interactive components while the agent retains control through intent messages — a design pattern for any agent that needs richer output than a text wall.
articleShopify Engineering
articleQwen Team
Qwen-Image: Crafting with Native Text Rendering
Qwen-Image is a 20B MMDiT image foundation model built for complex text rendering, including multi-line and paragraph-level layouts, alongside precise image editing.
Why it mattersQwen-Image is a 20B open-weights image model that finally handles complex in-image text rendering — multi-line layouts and both alphabetic and logographic scripts.
Unlocking the potential of vision language models on satellite imagery through fine-tuning
When prompts get long, brittle and inconsistent, adaptation is the better lever.
Why it mattersA concrete account of when LoRA fine-tuning beats prompt engineering for vision-language models: a handful of curated domain examples steers behavior more reliably than long, fragile prompts, with measured gains over base Pixtral-12B.
articlemistral.ai
Andrew Ng@AndrewYNg
The case that momentum, not current position, decides the AI race: China's open-weights ecosystem, hypercompetitive market and semiconductor push compound fast, while the US lead rests on advantages that policy alone won't sustain. Also a swipe at treating AGI as a finish line.
Why it mattersThe open-weights models many teams now depend on increasingly come from Chinese labs; this lays out why that momentum — and the semiconductor push behind it — is likely to continue rather than stall.
Announcing Codestral 25.08 and the Complete Mistral Coding Stack for Enterprise
A new Codestral version arrives alongside Mistral's argument about why AI coding tools stall in regulated enterprises.
Why it mattersCodestral 25.08 ships with a full coding stack aimed at VPC, on-prem and air-gapped environments, addressing the deployment and customization blockers that keep SaaS-only coding assistants out of regulated codebases.
articlemistral.ai
Gathering And Processing Raw Materials For The Model Factory
Every stage between a web crawl and a training batch, as Poolside runs it.
Why it mattersThe unglamorous half of model training, described concretely: how raw crawls and repos become Iceberg-backed assets, and why dependency sorting and packing matter as much as the filtering everyone talks about.
articlePoolside editorial sitemap
The Bitter Lesson versus The Garbage Can
An essay exploring whether AI agents can navigate messy, undocumented organizational processes by training on outcomes rather than encoding human expertise.
Why it mattersIf you're deploying agents inside real organizations, this reframes the problem: instead of mapping chaotic, undocumented workflows, you may only need to define good outputs and let agents learn the path.
articleEthan Mollick
articleQwen Team
GSPO: Towards Scalable Reinforcement Learning for Language Models
Scaling RL for language models requires stable training dynamics, and existing algorithms such as GRPO can collapse irreversibly over long runs.
Why it mattersGSPO stabilizes long-horizon RL training for language models by optimizing at the sequence level rather than per-token, directly addressing the model-collapse and instability failures teams hit when scaling GRPO.
Introducing Runway Aleph
Runway ships Aleph, an in-context video model that treats an existing clip as the conditioning signal.
Why it mattersAleph moves video generation from prompt-to-clip toward editing an existing shot: object removal, new camera angles and relighting run as tasks on input footage rather than regeneration from scratch.
articleRunway editorial sitemap
Multilingual Arbitrage Optimizing Data Pools To Accelerate Multilingual Progress 2024 08 28
Synthetic data pipelines usually pick one teacher model and stick with it.
Why it mattersGenerating synthetic data from one teacher model inherits its ceiling. Routing each language or slice to whichever model is strongest there lifts the resulting data pool without new human annotation.
articleCohere editorial sitemap
Aya Command 23 8b And 35b Technical Report 2024 05 23
Cohere Labs' technical report on the Aya 23 open-weight family, pairing the Command model backbone with the Aya instruction data and deliberately trading language breadth for depth on a smaller language set.
Why it mattersOpen-weight multilingual models at 8B and 35B give teams a self-hostable option for non-English workloads, with a published recipe explaining the depth-over-breadth language tradeoff.
articleCohere editorial sitemap
Include Evaluating Multilingual Language Understanding With Regional Knowledge 2024 11 29
Translated benchmarks measure whether a model can read a language, not whether it knows anything about where that language is spoken.
Why it mattersMost multilingual benchmarks are translated English, so they never test regional knowledge. INCLUDE sources questions from in-region exams instead, exposing gaps translation-based evals hide.
articleCohere editorial sitemap
Mix Data Or Merge Models Optimizing For Diverse Multi Task Learning 2024 10 15
Rather than assume a single blended training set, this work weighs data mixing against merging separately specialized checkpoints when a model must satisfy several objectives, including safety, at once.
Why it mattersIf you post-train for several objectives at once, this compares data mixing against model merging head to head and shows where merging avoids the tradeoffs that a single blended data pool forces.
articleCohere editorial sitemap
Llm See Llm Do Guiding Data Generation To Target Non Differentiable Objectives 2024 07 05
Synthetic data carries more than answers: the generating model passes its own characteristics into whatever is trained on its output, including ones nobody selected for.
Why it mattersTraining on another model's synthetic output silently inherits its properties, not just its capabilities, and the same channel can be used deliberately to steer objectives you cannot write a loss function for.
articleCohere editorial sitemap
Nexus Specialization Meets Adaptability For Efficiently Training Mixture Of Experts 2024 08 29
Nexus assembles a mixture of experts from independently trained dense models and adds a router designed to accept new experts after the fact, an attempt to keep specialization and the ability to adapt to unseen data distributions in the same architecture.
Why it mattersLets you build a mixture-of-experts model out of dense checkpoints you already trained, and extend it with new experts later instead of retraining the whole system when a new domain arrives.
articleCohere editorial sitemap
Pushing Mixture Of Experts To The Limit Extremely Parameter Efficient Moe For Instruction Tuning 2023 09 11
MoE usually implies training and serving a lot of parameters.
Why it mattersMixture-of-experts routing over lightweight adapters means you can instruction-tune while updating a fraction of a percent of parameters and still match or beat full fine-tuning at similar scale.
articleCohere editorial sitemap
Elo Uncovered Robustness And Best Practices In Language Model Evaluation 2023 11 29
A study of the Elo rating system as applied to language-model evaluation, probing whether its assumptions hold in the LLM setting and setting out practices that keep pairwise-comparison rankings reliable rather than an artifact of match ordering or sample size.
Why it mattersElo-style arenas drive a lot of model selection; this shows where those ratings are unstable and what sampling and ordering choices are needed before you trust a leaderboard gap.
articleCohere editorial sitemap
On The Challenges Of Using Black Box Apis For Toxicity Evaluation In Research 2023 04 24
Toxicity judgments shift with time, geography, and culture, and so do the commercial APIs that researchers lean on to measure them.
Why it mattersSafety evals built on a commercial classifier silently drift when the vendor updates it, so numbers reported months apart are not comparable. This is a trap for anyone tracking toxicity regressions over time.
articleCohere editorial sitemap
Rlhf Can Speak Many Languages Unlocking Multilingual Preference Optimization For Llms 2024 07 05
Preference optimization has become the default final training stage, yet its recipes are validated almost entirely in English.
Why it mattersPreference tuning recipes are tuned and validated in English by default. This shows how to extend RLHF-style optimization across languages and what data is needed for the gains to transfer.
articleCohere editorial sitemap
The Multilingual Alignment Prism Aligning Global And Local Preferences To Reduce Harm 2024 06 08
"Aligned" begs the question of aligned to whom.
Why it mattersSafety tuning on one region's preference data does not transfer cleanly; this measures how global versus local preference sources change harm rates when a model serves users in many languages.
articleCohere editorial sitemap
Aya Dataset Paper 2024 02 13
Cohere Labs documents the Aya collection.
Why it mattersAn openly licensed, human-curated multilingual instruction dataset gives teams a real alternative to machine-translated data when fine-tuning models for non-English users.
articleCohere editorial sitemap
Goodtriever Adaptive Toxicity Mitigation With Retrieval Augmented Models 2023 10 11
Goodtriever mitigates toxic generation with retrieval rather than parameter surgery.
Why it mattersSteering toxicity through a retrieval datastore rather than model weights means you can update what counts as harmful by editing data, with no retraining and no access to the parameters.
articleCohere editorial sitemap
One Tokenizer To Rule Them All Emergent Language Plasticity Via Multilingual Tokenizers 2025 05 30
Pretraining for many languages at once runs into capacity and data limits, so teams narrow scope early.
Why it mattersA tokenizer chosen at pretraining time constrains which languages you can adapt into later. Building it multilingual up front preserves that flexibility at little cost.
articleCohere editorial sitemap
Bam Just Like That Simple And Efficient Parameter Upcycling For Mixture Of Experts 2024 08 15
Turning a trained dense model into a mixture of experts usually copies only the feedforward layers.
Why it mattersMoE upcycling normally reuses only the dense model's feedforward layers. Carrying the attention parameters over too gives a better starting point for the same training budget.
articleCohere editorial sitemap
Lex Fridman
#475 – Demis Hassabis: Future of AI, Simulating Reality, Physics and Video Games
podcast
Listen
Flux 1 Kontext Dev
Black Forest Labs opened the weights of its Kontext editing model.
Why it mattersProprietary-grade instruction editing, including character consistency and local edits, becomes something you can run on your own hardware, with a paid path to ship it commercially.
articleBlack Forest Labs editorial sitemap
articleQwen Team
Qwen3-Coder: Agentic Coding in the World
Qwen3-Coder is Qwen's most agentic code model, led by a 480B mixture-of-experts variant with 35B active parameters, 256K native context and up to 1M tokens with extrapolation.
Why it mattersQwen3-Coder is an open-weight MoE model tuned specifically for agentic coding, tool-use, and browser-use, with 256K native (up to 1M) context and benchmark results rivaling Claude Sonnet 4.
Andrew Ng@AndrewYNg
With agentic coding compressing the cost of building to a spec, the constraint moves upstream to deciding what to build. Ng calls it the Product Management Bottleneck and argues teams need product intuition fast enough to keep pace with how quickly agents ship.
Why it mattersWhen implementation cost collapses, the queue of decisions becomes the bottleneck — this names the failure mode and argues for decision speed matched to how fast agents can now ship.
Compressing Context
Every multi-step agent run eventually outgrows its context window.
Why it mattersLong agent runs always exceed the window eventually; this lays out the retention, selection, and compression choices and how each one trades off answer quality against latency and token cost.
articlefactory.ai
The Big LLM Architecture Comparison
An in-depth technical article that compares the architectural designs of modern flagship open-weight LLMs, from DeepSeek-V3 and OLMo 2 to Kimi K2 and Gemma.
Why it mattersIf you're building on or evaluating open-weight LLMs, this gives you side-by-side architectural breakdowns (MLA vs GQA, MoE routing variants, normalization placement) with code.
articleSebastian Raschka, PhD
Le Chat dives deep.
Mistral adds a Deep Research mode to Le Chat that plans, searches and returns reference-backed reports.
Why it mattersLe Chat ships a planning-and-synthesis research agent, Voxtral voice input, and Magistral reasoning, forming a concrete alternative stack to evaluate against the assistants you already build on.
articlemistral.ai
Introducing The Model Factory
The overview of Poolside's internal training framework, organized around a claim.
Why it mattersFrames foundation model training as an automation problem rather than a staffing one, and enumerates the six capabilities a lab needs to iterate faster than it can hire.
articlePoolside editorial sitemap
Leveraging Multimodal Llms
How Shopify built its Global Catalogue.
Why it mattersA production blueprint for multimodal LLM extraction at billion-record scale: how to turn wildly inconsistent merchant-authored listings and images into standardized structured data that downstream agents can actually query.
articleShopify Engineering
Voxtral
Two open-weight speech models, a 24B for production scale and a 3B for edge and local deployment, combine transcription with understanding and multilingual coverage.
Why it mattersVoxtral ships speech understanding under Apache 2.0 at 24B and 3B, with a transcribe-optimized Mini variant served through the API. That gives developers open weights for local or edge ASR rather than a proprietary-only choice.
articlemistral.ai
Andrew Ng@AndrewYNg
Ng's AI Startup School talk on building fast with AI: why speed is the startup advantage, where opportunity sits across the AI stack, the shift toward agentic systems, and the specific execution habits AI Fund uses to shorten the loop from idea to running product.
Why it mattersA practitioner account of how a studio compresses the build-and-iterate loop with AI, including which layers of the stack are worth a startup's time and which are not.
Upgrading agentic coding capabilities with the new Devstral models
An upgrade tuned for scaffold generalization rather than benchmark fit.
Why it mattersDevstral Small 1.1 reaches 53.6% on SWE-Bench Verified under Apache 2.0 without test-time scaling, and Devstral Medium claims to beat Gemini 2.5 Pro and GPT-4.1 on code agents at a quarter of the cost.
articlemistral.ai
Andrew Ng@AndrewYNg
Agentic Document Extraction adds field-level extraction: point it at an invoice or medical form and get named fields back against a schema. If you'd rather not write the schema, one sample document plus a plain-language prompt generates it.
Why it mattersStructured field extraction from forms without hand-writing a JSON schema — upload one sample and describe what you want, and the schema is generated for you.
Andrew Ng@AndrewYNg
A walk through the three common post-training methods and what each one costs you in data: supervised fine-tuning on ideal outputs, DPO on chosen-versus-rejected pairs, and online RL on reward scores. Post-training is where a token predictor becomes an assistant.
Why it mattersLays out the practical difference between SFT, DPO and online RL by the data each demands, which is the decision most teams actually face when customizing a model on a budget.
Against "Brain Damage"
An essay by Ethan Mollick (Wharton Generative AI Lab) exploring how AI can help or hurt human thinking across learning, creativity, and teamwork.
Why it mattersA research-grounded rebuttal to the 'AI rots your brain' narrative, with concrete frameworks ('think first, write first, meet first') and shareable prompts for tutoring, ideation, and meeting facilitation.
articleEthan Mollick
Dwarkesh Patel
Why I don’t think AGI is right around the corner
podcast
Listen
LLM Research Papers: The 2025 List
A topic-organized collection of 200+ large language model research papers from the first half of 2025, curated by Sebastian Raschka.
Why it mattersA single, thematically organized reference to 200+ of the most important 2025 LLM papers — grouped by reasoning, RL, and multimodal themes.
articleSebastian Raschka, PhD
Agent Native Development
An article from Factory.ai arguing that AI agents can provide significant leverage for software development when used correctly.
Why it mattersLays out an opinionated methodology for structuring software work around agents, which is the core competence this audience is building.
articleFactory
Andrew Ng@AndrewYNg
A US district court ruled that training LLMs on copyrighted books is fair use, comparing the authors' complaint to objecting that schoolchildren learn to write. Read alongside the argument that data access, chips and regulatory capture are the real brakes on AI progress.
Why it mattersA US district court held that training LLMs on copyrighted books is fair use, reasoning by analogy to human learning — a data-access precedent for anyone training or fine-tuning on scraped text, appeal notwithstanding.
articleQwen Team
Qwen VLo: From "Understanding" the World to "Depicting" It
Qwen VLo moves the family from understanding images to generating them, unifying multimodal comprehension and generation in one model rather than treating them as separate systems.
Why it mattersQwen VLo unifies image understanding and generation in a single model, so you can prompt it to both interpret visual content and produce high-quality recreations without stitching together separate vision and image-gen models.
An index of the vibe-coding frontier. Corrections welcome.