Models2 min read
Reflection previews Beam, a 501B open-weight model still under evaluation
Early access is limited. The weights, model card, and technical report are due later this month.
AUTHOR
Kristian Kostov
About this portrait
Portrait adapted from Kristian’s photograph with AI.
Articles by Kristian
110 stories
Industry1 min read
ChatGPT will test visual ads during image generation
OpenAI plans a US test later this month and expands campaign measurement.
Research1 min read
OpenAI plans invisible text watermarks for ChatGPT and Codex in the EU
API customers can opt in now. OpenAI warns that detection has limits.
Tools & Agents1 min read
Cohere launches North 2 with reusable agents and spending controls
The enterprise platform adds memory, shared skills, and more deployment choices.
Industry1 min read
Cohere and PwC form an enterprise AI alliance, starting in Canada
The partnership targets enterprise search, research, and task automation.
Models2 min read
Google announces Gemini 4 Argon. Broad access is still coming.
Google has announced its next frontier model, with early access focused on trusted cyber defenders. Here is what builders can assess now.
Industry2 min read
Barclays plans Claude Code for half its developers by year’s end
The bank’s expanded Anthropic partnership offers a concrete view of how enterprise AI moves from pilots to everyday work.
Industry2 min read
Safeway brings grocery planning to ChatGPT. Checkout stays with the retailer.
Albertsons’ expanded OpenAI partnership connects conversational shopping with a familiar retail boundary: the final cart and purchase.
Research3 min read
Anthropic reports 6,157 disclosed bugs. The patching gap matters too.
An October dashboard update separates discovered candidates, disclosures, and known fixes. Those are different stages of security work.
Research2 min read
Google’s AlphaGenome Atlas makes billions of DNA predictions searchable
A public research resource turns a large prediction dataset into something researchers can explore. Predictions still need validation.
Models2 min read
GPT-6.1 Sol: OpenAI’s lower-cost model targets serious work
OpenAI’s new model puts capability and cost in the same conversation. Here’s what the announcement means for builders.
Industry2 min read
Anthropic commits $100M to train 10,000 AI engineers
Claude Frontier Academy aims to train 10,000 engineers by the end of 2027, with hands-on enterprise deployment at its core.
Models2 min read
Claude Sonnet 5.5: faster output, fewer tokens—same API rates
Anthropic says its latest Sonnet runs over 30% faster. The bigger question is what it costs to finish the job.
Tools & Agents2 min read
OpenAI’s dots bring always-on AI agents to everyday work
Always-on agents bring a new set of questions about access, responsibility, and how we measure useful work.
Research1 min read
Claude helped identify an enzyme system. Scientists verified it.
Anthropic’s biology research points to a promising discovery—and to the importance of experimental verification.
Tools & Agents2 min read
OpenAI DevDay 2026: cloud coding, plugins, and always-on agents
A look at the developer announcements, from cloud coding environments to tools for ongoing responsibilities.
Explainers2 min read
AI agents explained: what happens after you give them a goal?
A plain-language guide to tools, workflows, and systems that can choose their next step.
Explainers2 min read
RAG explained: how AI answers using your documents
How retrieval brings relevant documents into an answer, and why the evidence still needs a careful look.
Models3 min read
ChatGPT, Claude, or Gemini? Choose for the work you actually do
A practical comparison worksheet for your work, your sources, and your budget. No universal winner required.
Explainers3 min read
RAG vs fine-tuning: better knowledge or different behavior?
One supplies relevant information. The other adapts model behavior. Use the failure you see to choose the next step.
Tools & Agents3 min read
Before you trust an AI agent, check these five things
Check the result, permissions, and recovery. An agent saying “done” is only the beginning of verification.
Research2 min read
Your AI gave you a citation. Here’s how to check it.
A short claim-checking routine you can use before quoting, sharing, or publishing an AI answer.
Tools & Agents3 min read
MCP explained: how AI connects to tools and data
Understand how AI applications connect to tools and data, and what a connection does not guarantee.
Research3 min read
Why AI can sound certain—and still get the answer wrong
Why a convincing answer can be wrong, and how to make uncertainty visible before you rely on it.
Models3 min read
Your AI bill is more than tokens: tools, retries, and review
Use a worked example to estimate the cost of completed tasks, including the calls that do not succeed.
Explainers3 min read
Better AI prompts start with a clear brief
Give the task, source material, constraints, and output format. Then test whether the answer meets the brief.
Models1 min read
DeepSeek-V3 explained: architecture, strengths, and limits
DeepSeek-V3 is a mixture-of-experts language model described in DeepSeek’s 2024 technical report.
Models1 min read
DeepSeek-R1 explained: reasoning models and distilled versions
DeepSeek-R1 explores reasoning training with reinforcement learning.
Models1 min read
Qwen2.5 explained: choosing a model within the family
Qwen2.5 is a family of language models documented by the Qwen team.
Models1 min read
Qwen3 explained: dense models, experts, and thinking modes
The Qwen3 technical report describes dense and mixture-of-experts models, with support for thinking and non-thinking behavior.
Models1 min read
Llama 3 explained: what the technical report covers
Meta’s Llama 3 report describes a family of foundation models and instruction-tuned variants.
Models1 min read
Mistral 7B explained: why a smaller model can be useful
Mistral 7B is the language model described in Mistral’s 2023 paper.
Models1 min read
Mixtral explained: total parameters versus active experts
Mixtral uses a sparse mixture-of-experts architecture.
Models1 min read
Gemma 2 explained: model size and deployment tradeoffs
Gemma 2 is the open-weight model generation described in Google’s technical report.
Models1 min read
Phi-3 explained: what a small language model can do
Microsoft’s Phi-3 report studies compact language models trained with an emphasis on data quality.
Models1 min read
OLMo 2 explained: why training transparency matters
OLMo 2 is a language-model family documented by the Allen Institute for AI and collaborators.
Models1 min read
SmolLM2 explained: a compact model for local experiments
SmolLM2 is a compact language-model family from Hugging Face.
Models1 min read
Whisper explained: speech recognition is not fact checking
Whisper is a speech-recognition system described in OpenAI’s research on weakly supervised audio training.
Models1 min read
CLIP explained: connecting images with text descriptions
CLIP learns a relationship between images and text through paired training examples.
Models1 min read
Sentence-BERT explained: a model for text similarity
Sentence-BERT adapts BERT-style models to produce useful sentence embeddings.
Models1 min read
BERT explained: an encoder rather than a chat assistant
BERT stands for Bidirectional Encoder Representations from Transformers.
Models1 min read
T5 explained: treating language tasks as text-to-text problems
T5 explores a unified text-to-text approach to language tasks.
Tools & Agents1 min read
AI agents vs workflows: which should you build?
A workflow follows predefined steps; an agent chooses how to proceed using feedback and tools.
Tools & Agents1 min read
Prompt chaining: split an AI task into checkable steps
Prompt chaining passes the output of one model call into another.
Tools & Agents1 min read
Model routing: send different tasks to different systems
Routing classifies a request and sends it to a suitable processing path.
Tools & Agents1 min read
Parallel AI workflows: when independent tasks can run together
Parallel workflows run independent subtasks at the same time and combine their results.
Tools & Agents1 min read
The evaluator–optimizer pattern: revise against a clear standard
An evaluator–optimizer workflow separates generation from feedback and revision.
Tools & Agents1 min read
Design AI tools that an agent can use correctly
An agent tool needs a clear purpose, precise inputs, and a result the model can interpret.
Tools & Agents1 min read
AI tool errors: recover without repeating unsafe actions
A tool error should tell the agent what failed and whether an action completed.
Tools & Agents1 min read
Agent sandboxing: limit what automated code can reach
A sandbox restricts the environment in which an agent or its code runs.
Tools & Agents1 min read
What should an AI agent remember—and what should it forget?
Agent memory preserves information beyond the current model context.
Tools & Agents1 min read
Context compaction: shorten history without losing the task
Context compaction condenses a long working history into the information needed to continue.
Tools & Agents1 min read
MCP tools vs resources: actions and context are different
MCP resources provide contextual data, while tools expose executable functions.
Tools & Agents1 min read
MCP prompts: reusable instructions, with boundaries
MCP prompts expose reusable message templates through a server.
Tools & Agents1 min read
AI that clicks your screen: why each next step matters
A computer-use agent acts through an interface and must observe the resulting state.
Tools & Agents1 min read
Multi-agent systems: account for coordination as well as work
Multiple agents can divide work, but they also add communication, duplicate context, and integration overhead.
Tools & Agents1 min read
Where to put the human review in an AI workflow
An approval gate gives a person a concrete result to review before a consequential tool action.
Tools & Agents1 min read
Build an agent evaluation suite before granting more autonomy
An agent evaluation suite tests complete task outcomes, not only the final text.
Research1 min read
Attention Is All You Need: what the transformer paper changed
The 2017 transformer paper introduced an attention-based architecture without the recurrent or convolutional components used in many earlier sequence models.
Research1 min read
Chinchilla explained: model size and training data must be balanced
The Chinchilla research examines model size and training-token allocation under a fixed compute budget.
Research1 min read
RLHF explained through the InstructGPT research
Reinforcement learning from human feedback uses preference information to improve model behavior.
Research1 min read
Direct Preference Optimization: a simpler route to preferences
Direct Preference Optimization, or DPO, trains a language model using preferred and rejected responses without the same explicit reinforcement-learning pipeline used in conventional RLHF.
Research1 min read
Constitutional AI: using principles to guide model feedback
Constitutional AI studies using a set of principles to guide critique, revision, and AI-generated feedback.
Research1 min read
LoRA research explained: adapt a model with fewer trainable weights
LoRA freezes the pretrained model weights and trains smaller low-rank updates.
Research1 min read
QLoRA explained: combine quantization with adapter training
QLoRA combines a quantized frozen base model with trainable low-rank adapters.
Research1 min read
Lost in the Middle: why a long context can still miss the answer
Lost in the Middle studies how models use relevant information placed at different positions in long inputs.
Research1 min read
Self-consistency research: several answers can still share a mistake
Self-consistency samples multiple reasoning paths and combines their final answers.
Research1 min read
SWE-bench explained: repository tasks rather than isolated snippets
SWE-bench evaluates software tasks based on real repository issues and associated fixes.
Research1 min read
HumanEval and pass@k: what a coding score really measures
HumanEval is a set of programming problems introduced with research on code-trained language models.
Research1 min read
HELM explained: evaluate more than accuracy
HELM stands for Holistic Evaluation of Language Models.
Research1 min read
TruthfulQA explained: fluent answers can repeat misconceptions
TruthfulQA evaluates whether models produce truthful answers to questions designed around common misconceptions.
Research1 min read
ReAct explained: combine reasoning with external observations
ReAct combines reasoning and actions in a task-solving process.
Research1 min read
Toolformer explained: learning when a language model should call tools
Toolformer studies teaching a model when and how to use external tools, including interpreting their outputs.
Research1 min read
FlashAttention explained: faster attention without approximating it
FlashAttention is an attention algorithm designed around GPU memory movement.
Research1 min read
Speculative decoding: draft tokens, then verify them
Speculative decoding uses a faster draft process to propose tokens, then checks them with a target model.
Industry1 min read
How to evaluate an AI vendor beyond the demo
Evaluate an AI vendor against your actual workflow, data requirements, and failure tolerance.
Industry1 min read
Is your AI pilot saving time—or moving work into review?
An AI pilot should measure useful work after review, not just generated output.
Industry1 min read
The AI costs that token prices leave out
Total cost of ownership includes operating the complete AI system.
Industry1 min read
What makes an AI provider hard to leave?
Provider lock-in can come from APIs, prompt formats, tools, stored state, and operational dependencies.
Industry1 min read
Open weights is not the same as open source AI
Open weights means model parameters are available under some terms.
Industry1 min read
A practical checklist for reviewing a model’s license
A model’s license governs permitted use under its stated terms.
Industry1 min read
AI procurement: questions to ask before signing up
AI procurement should establish what a service can do, what data it handles, and how failures will be managed.
Industry1 min read
AI API data retention: check the specific feature, not the slogan
Data-retention rules can differ by product, model, feature, and account configuration.
Industry1 min read
AI data residency: map the request before choosing a region
Data residency concerns where data is stored or processed under a service’s specific commitments.
Industry1 min read
Model deprecation: build a migration path before an endpoint disappears
Hosted models have versions and lifecycle policies.
Industry1 min read
AI governance: assign an owner to the system and its failures
AI governance establishes responsibility for decisions, monitoring, and response.
Industry1 min read
How to read an AI company announcement without losing the qualifiers
An AI announcement combines reported facts, vendor measurements, and positioning.
Industry1 min read
AI data centres and electricity: read projections as scenarios
AI depends on electricity for data-centre computing.
Industry1 min read
AI chips: bandwidth, memory, and software matter alongside compute
AI accelerators perform specialized computations used in model training and inference.
Industry1 min read
Managed AI hosting vs self-hosting: choose the operating model
Managed hosting delegates parts of model-serving operations to a provider; self-hosting gives your team more operational control.
Industry1 min read
AI skills for teams: teach verification before speed
A useful AI training program teaches people to frame tasks, inspect sources, protect data, and verify outcomes.
Industry1 min read
How to read a model card before using a checkpoint
A model card documents information about a model, such as its intended use, training context, evaluations, and limitations.
Industry1 min read
That AI benchmark score needs a closer look
A benchmark chart is meaningful only with its task set, settings, and scoring method.
Industry1 min read
Stealth AI models: a mysterious name is not proof of a breakthrough
An anonymous model label hides the provider’s identity during testing.
Explainers1 min read
What are AI tokens, and why are they not the same as words?
Tokens are the units a model receives and generates.
Explainers1 min read
What are text embeddings? A short guide to meaning as vectors
A text embedding represents text as a numerical vector.
Explainers1 min read
Vector search explained: nearest does not always mean relevant
Vector search finds stored vectors close to a query vector under a chosen distance or similarity measure.
Explainers1 min read
A bigger context window does not guarantee a better answer
A context window limits the information a model can work with during a request.
Explainers1 min read
AI temperature explained: a sampling setting, not an accuracy dial
Temperature changes the probabilities used when sampling a model’s next token.
Explainers1 min read
Top-p explained: limit the candidate tokens before sampling
Top-p, also called nucleus sampling, keeps a set of likely next tokens whose combined probability reaches a threshold.
Explainers1 min read
Training vs inference: learning weights and using a model
Training adjusts model parameters using data and an optimization process.
Explainers1 min read
When examples should change how your model behaves
Fine-tuning continues training a pretrained model on a selected dataset.
Explainers1 min read
Smaller AI models, less memory: what quantization trades away
Quantization stores or computes model values at reduced numerical precision.
Explainers1 min read
Mixture of experts explained: many specialists, selective work
A mixture-of-experts model contains multiple expert components and a routing mechanism.
Explainers1 min read
Multimodal AI explained: a model can accept more than text
Multimodal systems work with more than one type of input or output, such as text and images.
Explainers1 min read
Reading the text is not the same as understanding the document
OCR extracts text from an image or scanned document.
Explainers1 min read
When a document tries to give your AI instructions
Prompt injection attempts to steer a model through untrusted input, such as a document or tool result.
Explainers1 min read
Structured AI output: valid JSON is only the first check
Structured output constrains a response to a specified format or schema.
Explainers1 min read
Streaming AI responses: show progress before the answer is complete
Streaming delivers parts of a generated response as they become available.
Explainers1 min read
Batch AI processing: use asynchronous jobs for non-urgent work
Batch processing groups requests into an asynchronous job.