All Posts
ChatGPT Work pushed desktop agents toward the default
By: Jeremy London
Published: Jul 12, 2026 // Category: agent-workflows
Description: By July 2026, the agent conversation had moved from chat help toward long-running work across apps and files.Read more
Fable 5, Sonnet 5, and GPT-5.6 raised the safety and routing bar
By: Jeremy London
Published: Jul 10, 2026 // Category: model-watch
Description: The 2026 frontier launches made raw capability harder to separate from safeguards, effort levels, pricing, and model routing.Read more
Being recognized as AI Security Innovator of the Year
By: Jeremy London
Published: Jul 1, 2026 // Category: security-trust
Description: The Hacker News Cybersecurity Stars Awards recognized me for the AI security work I have been building at Keeper.Read more
Winning the Opir Challenge by testing the edges
By: Jeremy London
Published: Jun 29, 2026 // Category: security-trust
Description: I was one of the first-place winners of Wordcab's Opir Challenge after probing where semantic guardrails get thin.Read more
Portfolios age in public
By: Jeremy London
Published: Jun 27, 2026 // Category: personal-systems
Description: A personal site needs occasional pruning because old work keeps speaking after the context disappears.Read more
Codex sandboxing is the agent story I care about
By: Jeremy London
Published: Jun 24, 2026 // Category: security-trust
Description: The interesting agent work in 2026 is less about personality and more about controlled file access, network limits, and recovery.Read more
Private AI workflows want local state
By: Jeremy London
Published: Jun 18, 2026 // Category: local-first-software
Description: Local state matters more when AI tools start reading notes, files, and drafts that were never meant for a server round trip.Read more
AI spend caps belong in product design
By: Jeremy London
Published: Jun 12, 2026 // Category: security-trust
Description: As AI platforms added spend controls, it became clearer that budgets are part of user experience.Read more
On-device AI fits private workflows
By: Jeremy London
Published: Jun 8, 2026 // Category: deep-learning-networks
Description: On-device AI brings models closer to the user, which matters for privacy, latency, offline use, and personal tooling.Read more
Small sites need smoke tests too
By: Jeremy London
Published: Jun 1, 2026 // Category: web-engineering
Description: A personal site can still break routes, feeds, images, and content collections if nobody checks the full path.Read more
Opir made guardrails look like classifiers again
By: Jeremy London
Published: May 29, 2026 // Category: model-watch
Description: Opir put safety filtering back in the small-model conversation with encoder guardrails for toxicity, jailbreaks, and policy categories.Read more
Dependencies are part of the threat model
By: Jeremy London
Published: May 28, 2026 // Category: security-trust
Description: Package choices affect security, maintenance, performance, and how quickly a team can patch.Read more
Fine-tuning is not a shortcut
By: Jeremy London
Published: May 19, 2026 // Category: advance-deep-learning
Description: Fine-tuning can help, but it works best after the task, data, evals, and product behavior are already clear.Read more
Node 26 reminded me runtime upgrades are calendar work
By: Jeremy London
Published: May 10, 2026 // Category: developer-tools
Description: Node release lines keep moving, and the least painful teams treat runtime upgrades as scheduled maintenance.Read more
AI-generated UI still needs a design system
By: Jeremy London
Published: May 6, 2026 // Category: design-systems
Description: AI can generate interface pieces quickly, which makes design-system boundaries more important rather than less.Read more
Model routing is a systems problem
By: Jeremy London
Published: Apr 24, 2026 // Category: advance-deep-learning
Description: Model routing is about choosing the right model for the job while balancing quality, speed, cost, policy, and risk.Read more
AI incident reviews need model context
By: Jeremy London
Published: Apr 22, 2026 // Category: ai-operations
Description: An AI incident review has to include prompt, model, retrieved context, tool calls, and approval state.Read more
Fresh context review is not bureaucracy
By: Jeremy London
Published: Apr 18, 2026 // Category: agent-workflows
Description: Verifier agents catch different failures when they are not soaked in the builder rationale.Read more
SWE-bench stopped being enough
By: Jeremy London
Published: Apr 8, 2026 // Category: developer-tools
Description: When a benchmark stops separating frontier coding systems, teams need local evals more, not less.Read more
Tooling should make the safe path short
By: Jeremy London
Published: Mar 22, 2026 // Category: developer-tools
Description: If the careful workflow takes five commands and the risky workflow takes one, the risky workflow wins.Read more
Product specs should name the boring states
By: Jeremy London
Published: Mar 18, 2026 // Category: product-engineering
Description: A product spec that skips loading, empty, error, permission, and rollback states is not finished.Read more
Multimodal embeddings made search feel wider
By: Jeremy London
Published: Mar 14, 2026 // Category: advance-deep-learning
Description: By 2026, embedding work was moving beyond text into image, audio, video, and PDF retrieval.Read more
Eval dashboards for real feedback
By: Jeremy London
Published: Mar 10, 2026 // Category: advance-deep-learning
Description: An eval dashboard turns model examples into a clearer view of quality, regressions, latency, and cost.Read more
Tool calling is a contract
By: Jeremy London
Published: Feb 27, 2026 // Category: deep-learning-principals
Description: Tool calling works best when the model, tool schema, identity, and application agree on a clear contract.Read more
AI visual explainers need respect for the reader
By: Jeremy London
Published: Feb 25, 2026 // Category: creative-coding
Description: AI-generated visuals can help technical writing, but only when the visual teaches something the text cannot.Read more
Model retirements made AI maintenance visible
By: Jeremy London
Published: Feb 16, 2026 // Category: deep-learning-principals
Description: By early 2026, model menus and deprecation dates made it obvious that AI applications need maintenance plans.Read more
Technical writing needs room for wrong guesses
By: Jeremy London
Published: Feb 11, 2026 // Category: learning-lab
Description: A good tutorial shows the mistake the reader is about to make and gives them a way back.Read more
Audit logs should answer human questions
By: Jeremy London
Published: Jan 30, 2026 // Category: security-trust
Description: Logs should tell a person what happened, who did it, and what changed without a forensic treasure hunt.Read more
Small data warehouses need maintenance too
By: Jeremy London
Published: Jan 21, 2026 // Category: data-engineering
Description: Even small analytical stores need ownership, naming, retention, and boring checks before they feed product decisions.Read more
Multimodal systems need grounded workflows
By: Jeremy London
Published: Jan 15, 2026 // Category: advance-deep-learning
Description: Multimodal models are useful when they fit into a real workflow with clear inputs, checks, and ownership.Read more
Smaller models belong in the architecture
By: Jeremy London
Published: Dec 30, 2025 // Category: deep-learning-networks
Description: A practical note on using smaller models for focused jobs instead of sending every request to the largest model.Read more
Maintenance days are real work
By: Jeremy London
Published: Dec 27, 2025 // Category: personal-systems
Description: Cleaning dependencies, notes, drafts, and small broken workflows is work, even when it does not feel like shipping.Read more
AI SDK 6 made tool approval mainstream
By: Jeremy London
Published: Dec 24, 2025 // Category: agent-workflows
Description: AI SDK 6 put agents, MCP, tool approval, and DevTools in the same conversation, which is where production AI apps were heading.Read more
Local dev needs an exit ramp
By: Jeremy London
Published: Dec 9, 2025 // Category: developer-tools
Description: A development loop should tell you how to stop, reset, and recover without reading every script.Read more
Sonnet 4.5 and Gemini 3 made coding the frontier battleground
By: Jeremy London
Published: Nov 19, 2025 // Category: model-watch
Description: Anthropic and Google ended 2025 by treating coding, agents, and computer use as the proving ground for frontier models.Read more
RSS is still good plumbing
By: Jeremy London
Published: Nov 15, 2025 // Category: web-engineering
Description: Feeds are unglamorous, durable, and still worth maintaining for a personal site.Read more
Synthetic data is useful for evals
By: Jeremy London
Published: Nov 8, 2025 // Category: advance-deep-learning
Description: Synthetic data can help teams test rare cases, but it should not replace real examples from production.Read more
Guardrails are product decisions
By: Jeremy London
Published: Oct 21, 2025 // Category: advance-deep-learning
Description: AI guardrails define what a product should do, refuse, explain, log, and escalate.Read more
Codex GA made agent admin a real job
By: Jeremy London
Published: Oct 8, 2025 // Category: security-trust
Description: When coding agents move from preview into regular team use, admin controls become part of engineering quality.Read more
Math clicks when the shapes stop moving
By: Jeremy London
Published: Oct 3, 2025 // Category: learning-lab
Description: Stable notation and repeated small examples do more for intuition than a tour through every formula.Read more
Semantic search needs permissions
By: Jeremy London
Published: Sep 16, 2025 // Category: advance-deep-learning
Description: Semantic search is useful, but production systems need metadata, access control, and clear retrieval behavior.Read more
Approval buttons need context
By: Jeremy London
Published: Sep 2, 2025 // Category: agent-workflows
Description: An approval step only helps if the human can see the plan, the data, and the rollback story.Read more
Queues make AI workflows calmer
By: Jeremy London
Published: Aug 29, 2025 // Category: deep-learning-principals
Description: Background queues help AI platforms handle slow jobs, retries, bursts, and long-running agent workflows.Read more
Claude 4 and GPT-5 made agent work a maintenance problem
By: Jeremy London
Published: Aug 8, 2025 // Category: model-watch
Description: Claude 4 pushed long-running coding agents, then GPT-5 turned automatic reasoning into the default ChatGPT path.Read more
Agent observability needs user-level events
By: Jeremy London
Published: Aug 7, 2025 // Category: ai-operations
Description: AI operations need logs that explain what the agent did in product terms, not only infrastructure terms.Read more
Secret scans are cheap compared to rotating keys
By: Jeremy London
Published: Jul 28, 2025 // Category: security-trust
Description: A secret scanner is boring right up until it saves a weekend.Read more
DeepSeek, Kimi, and MiniMax made open reasoning cheaper
By: Jeremy London
Published: Jul 15, 2025 // Category: model-watch
Description: The 2025 Chinese open-weight wave changed the cost story around reasoning, coding, and agentic model work.Read more
AI SDK 5 made chat state an interface
By: Jeremy London
Published: Jul 14, 2025 // Category: web-engineering
Description: Vercel AI SDK 5 treated chat as a typed, customizable application surface instead of a streaming demo.Read more
Caching AI responses with care
By: Jeremy London
Published: Jul 12, 2025 // Category: deep-learning-principals
Description: Caching can make AI apps faster and cheaper, but production systems need careful keys, freshness rules, and safety checks.Read more
Ranking shows up everywhere
By: Jeremy London
Published: Jun 25, 2025 // Category: deep-learning-networks
Description: Ranking shows up in recommendations, retrieval, alerts, model routing, and engineering workflows.Read more
Sync conflicts are interface problems
By: Jeremy London
Published: Jun 18, 2025 // Category: local-first-software
Description: Conflict resolution belongs in the product experience as well as the replication layer.Read more
Automation should leave a receipt
By: Jeremy London
Published: Jun 14, 2025 // Category: personal-systems
Description: Personal automation is easier to trust when it leaves a plain record of what it did and why.Read more
The first example carries too much weight
By: Jeremy London
Published: May 31, 2025 // Category: learning-lab
Description: Readers build their mental model from the first runnable example, so that example has to be honest.Read more
Claude 4 made long-running agent work a planning problem
By: Jeremy London
Published: May 27, 2025 // Category: agent-workflows
Description: Claude 4 pushed the conversation toward sustained coding and agent workflows, which raises the bar for supervision.Read more
Codex made parallel agent work feel normal
By: Jeremy London
Published: May 20, 2025 // Category: agent-workflows
Description: The Codex research preview made it easier to imagine software agents working in separate sandboxes instead of one chat thread.Read more
Threat analytics is time series work too
By: Jeremy London
Published: May 7, 2025 // Category: deep-learning-networks
Description: Security signals often come down to timing, baselines, drift, and knowing when a pattern changed.Read more
Qwen 2.5 and Qwen3 made open models feel global
By: Jeremy London
Published: Apr 30, 2025 // Category: model-watch
Description: Alibaba's Qwen releases showed that the open-weight race was no longer centered on one country, one lab, or one model shape.Read more
o3 and o4-mini made routing harder to dodge
By: Jeremy London
Published: Apr 23, 2025 // Category: deep-learning-networks
Description: The o-series releases made the cheap, fast, smart tradeoff harder to avoid and better to name.Read more
Generative systems need constraints
By: Jeremy London
Published: Apr 18, 2025 // Category: advance-deep-learning
Description: Generative AI is useful when the system has constraints, review paths, and clear expectations for output quality.Read more
Spring 2025 split reasoning into products
By: Jeremy London
Published: Apr 18, 2025 // Category: model-watch
Description: Claude 3.7 Sonnet, Gemini 2.5 Pro, GPT-4.1, o3, and o4-mini made reasoning less like one capability and more like a set of product choices.Read more
GPT-4.1 made coding models feel specialized
By: Jeremy London
Published: Apr 16, 2025 // Category: developer-tools
Description: The GPT-4.1 API release made it clearer that coding performance is a product choice, not one generic model score.Read more
Shader toys are debugging tools too
By: Jeremy London
Published: Apr 9, 2025 // Category: creative-coding
Description: Small visual programs can be toys and debugging tools because they make math and state visible.Read more
Forms deserve the boring states
By: Jeremy London
Published: Apr 5, 2025 // Category: web-engineering
Description: The quality of a form shows up in loading, empty, invalid, and submitted states more than in the first screenshot.Read more
Gemini 2.5 made long context feel operational
By: Jeremy London
Published: Mar 31, 2025 // Category: advance-deep-learning
Description: Gemini 2.5 Pro pushed long-context reasoning further into ordinary developer planning and product work.Read more
Git history should explain the decision
By: Jeremy London
Published: Mar 19, 2025 // Category: developer-tools
Description: A commit message is one of the cheapest places to preserve why a change happened.Read more
Data contracts are where AI quality starts
By: Jeremy London
Published: Mar 6, 2025 // Category: data-engineering
Description: Model quality work starts before the prompt when the data shape is undocumented or quietly changing.Read more
Small model choices still matter
By: Jeremy London
Published: Mar 5, 2025 // Category: deep-learning-basics
Description: A short note on the small model details that still shape behavior when systems move into production.Read more
Claude 3.7 made thinking a product control
By: Jeremy London
Published: Feb 28, 2025 // Category: deep-learning-networks
Description: Claude 3.7 Sonnet made hybrid reasoning visible enough that teams had to decide when thinking should cost time.Read more
Notes need friction in the right place
By: Jeremy London
Published: Feb 22, 2025 // Category: personal-systems
Description: Fast capture is useful, but the real value comes from making review and reuse easy enough to happen.Read more
Data cleaning is model work
By: Jeremy London
Published: Feb 14, 2025 // Category: deep-learning-principals
Description: Better data often improves AI systems more than another round of prompt tweaks.Read more
Forms are where product quality leaks
By: Jeremy London
Published: Feb 5, 2025 // Category: product-engineering
Description: Form quality shows up in pending, failed, duplicate, and partially valid states more than in the first screenshot.Read more
Input validation is product design
By: Jeremy London
Published: Feb 3, 2025 // Category: security-trust
Description: Validation rules are defensive code, but they also describe what the product accepts as real.Read more
Tailwind 4 made tokens feel closer to CSS
By: Jeremy London
Published: Jan 30, 2025 // Category: design-systems
Description: Tailwind 4 made the styling system feel more native to CSS, which is a good excuse to clean up theme decisions.Read more
Tailwind 4 made CSS configuration feel like CSS
By: Jeremy London
Published: Jan 29, 2025 // Category: web-engineering
Description: Tailwind CSS v4 moved more of the system into CSS itself, which made configuration feel less separate from styling.Read more
AI apps need boring observability
By: Jeremy London
Published: Jan 23, 2025 // Category: advance-deep-learning
Description: Production AI needs logs, traces, prompt records, retrieval records, and examples that can become evals.Read more
Handoffs are part of the agent
By: Jeremy London
Published: Jan 16, 2025 // Category: agent-workflows
Description: A handoff file is not ceremony. It is how an agent leaves enough state for the next run to be useful.Read more
Pyodide makes blog posts feel less theoretical
By: Jeremy London
Published: Jan 5, 2025 // Category: learning-lab
Description: Running Python in the page changes a tutorial from explanation into a small lab.Read more
o1 and Gemini 2.0 pushed agents toward the default
By: Jeremy London
Published: Dec 12, 2024 // Category: model-watch
Description: OpenAI's o1 and Google's Gemini 2.0 moved the conversation from chat quality to reasoning time, tools, and agent-shaped product paths.Read more
Prompt pipelines need ownership
By: Jeremy London
Published: Dec 11, 2024 // Category: deep-learning-principals
Description: Prompts become more reliable when teams treat them like versioned software with tests, owners, and review.Read more
React 19 made form state feel native again
By: Jeremy London
Published: Dec 10, 2024 // Category: web-engineering
Description: React 19 stabilized a set of APIs that made async UI feel less like plumbing and more like application code.Read more
CSS gets better when it has fewer opinions
By: Jeremy London
Published: Dec 4, 2024 // Category: web-engineering
Description: Most site styling improves when the design system has fewer special cases and more predictable defaults.Read more
MCP turned agent tools into an interface
By: Jeremy London
Published: Dec 2, 2024 // Category: agent-workflows
Description: The Model Context Protocol gave agent tool access a shared shape, which made the design questions clearer.Read more
Local models are part of the platform
By: Jeremy London
Published: Nov 19, 2024 // Category: deep-learning-networks
Description: Local inference is useful for privacy, latency, and focused workflows, especially when it is treated as part of the system.Read more
Scripts should be idempotent by default
By: Jeremy London
Published: Nov 7, 2024 // Category: developer-tools
Description: The best project scripts can run twice without making the second run weird.Read more
Next 15 made upgrades feel like product work
By: Jeremy London
Published: Oct 24, 2024 // Category: web-engineering
Description: Next.js 15 was a reminder that framework upgrades are user-facing when routing and request APIs change.Read more
Authorization is where apps get personal
By: Jeremy London
Published: Oct 24, 2024 // Category: security-trust
Description: Authentication says who showed up. Authorization decides whether the app respects what belongs to them.Read more
Agents need boring boundaries
By: Jeremy London
Published: Oct 6, 2024 // Category: advance-deep-learning
Description: A practical note on agents, MCP tools, identity, permissions, and why clear limits matter.Read more
Postgres 17 reminded me JSON is table work too
By: Jeremy London
Published: Oct 2, 2024 // Category: data-engineering
Description: Postgres 17 put more SQL/JSON work into the database, which made messy application data feel less separate from relational design.Read more
LFM made small models an architecture question
By: Jeremy London
Published: Oct 1, 2024 // Category: model-watch
Description: Liquid Foundation Models made small language models feel less like compressed LLMs and more like a different design bet.Read more
RAG only works if retrieval is boringly good
By: Jeremy London
Published: Sep 28, 2024 // Category: advance-deep-learning
Description: Retrieval augmented generation depends less on a clever prompt and more on reliably finding the right context.Read more
A writing backlog should not be a guilt machine
By: Jeremy London
Published: Sep 19, 2024 // Category: personal-systems
Description: A useful backlog stores ideas without turning every old note into a quiet accusation.Read more
Reasoning models changed where latency belongs
By: Jeremy London
Published: Sep 18, 2024 // Category: deep-learning-networks
Description: The o1 preview made it easier to talk about spending extra time on hard steps instead of every step.Read more
Why I like small agent tools
By: Jeremy London
Published: Sep 3, 2024 // Category: agent-workflows
Description: Small tools are easier to permission, easier to review, and harder for an agent to misuse quietly.Read more
Visual examples should earn their space
By: Jeremy London
Published: Aug 29, 2024 // Category: learning-lab
Description: A diagram helps when it lets the learner predict the next step, not when it decorates the page.Read more
Interactive essays need state, not screenshots
By: Jeremy London
Published: Aug 21, 2024 // Category: creative-coding
Description: An interactive technical essay should let the reader change the system instead of only admiring a captured result.Read more
Evals are how AI teams stay honest
By: Jeremy London
Published: Aug 13, 2024 // Category: deep-learning-principals
Description: A simple note on using evals to catch regressions before users find them.Read more
Structured Outputs made prompts less magical
By: Jeremy London
Published: Aug 12, 2024 // Category: deep-learning-principals
Description: OpenAI Structured Outputs moved a common integration problem from prompt superstition into schema work.Read more
Astro content is a real interface
By: Jeremy London
Published: Aug 11, 2024 // Category: web-engineering
Description: Static content systems still need schema discipline, naming conventions, and boring build checks.Read more
Good CLI output is a kind of API
By: Jeremy London
Published: Jul 30, 2024 // Category: developer-tools
Description: Terminal output should be written for the person debugging at 11 p.m., not for the happy path demo.Read more
Llama 3 and 3.1 made open weight the default baseline
By: Jeremy London
Published: Jul 24, 2024 // Category: model-watch
Description: Meta's 2024 Llama releases changed the question from whether to test open weights to where they fit in the stack.Read more
Vector stores need product discipline
By: Jeremy London
Published: Jul 22, 2024 // Category: advance-deep-learning
Description: Vector databases help AI systems remember, but production quality depends on chunks, metadata, permissions, and evals.Read more
Agent runbooks need failure examples
By: Jeremy London
Published: Jul 21, 2024 // Category: agent-workflows
Description: Runbooks for agents get better when they include the mistakes the agent is expected to make.Read more
TypeScript 5.5 cut some friction from filter code
By: Jeremy London
Published: Jul 10, 2024 // Category: developer-tools
Description: Inferred type predicates were not flashy, but they removed a small daily tax from TypeScript code.Read more
Claude 3.5 Sonnet raised the baseline for workflow tools
By: Jeremy London
Published: Jun 25, 2024 // Category: agent-workflows
Description: Claude 3.5 Sonnet made many tool workflows feel less fragile, which also made weak process harder to excuse.Read more
GPT-4o and Claude 3.5 Sonnet compressed the frontier
By: Jeremy London
Published: Jun 22, 2024 // Category: model-watch
Description: The May and June 2024 launches made frontier quality feel faster, cheaper, more visual, and harder to rank with one benchmark.Read more
DuckDB made local analytics feel normal
By: Jeremy London
Published: Jun 12, 2024 // Category: data-engineering
Description: DuckDB 1.0 made it easier to treat local analytical work as real engineering instead of a notebook side quest.Read more
Embeddings are product plumbing
By: Jeremy London
Published: Jun 9, 2024 // Category: deep-learning-principals
Description: Embeddings are useful because they make search, memory, deduping, and retrieval feel less brittle.Read more
GPT-4o made multimodal product work real
By: Jeremy London
Published: May 20, 2024 // Category: advance-deep-learning
Description: The GPT-4o launch made audio, vision, and text feel like one product surface instead of three demos.Read more
Routing is the quiet part of AI platforms
By: Jeremy London
Published: May 17, 2024 // Category: deep-learning-networks
Description: A short note on model routing, context routing, and why production AI systems need more than one path.Read more
GLiNER turned entity extraction into a small-model job
By: Jeremy London
Published: May 10, 2024 // Category: model-watch
Description: GLiNER was a useful reminder that not every language task needs a decoder LLM pretending to be a classifier.Read more
SQLite in the browser changes the default
By: Jeremy London
Published: May 8, 2024 // Category: local-first-software
Description: When local databases become easy to ship, small tools can keep more state close to the user.Read more
Kubernetes 1.30 was an upgrade planning reminder
By: Jeremy London
Published: Apr 24, 2024 // Category: ai-operations
Description: Kubernetes releases are a calendar discipline that reaches beyond cluster administration.Read more
Gemma and Phi made small models feel serious again
By: Jeremy London
Published: Apr 23, 2024 // Category: model-watch
Description: Google Gemma and Microsoft Phi-3 pushed small open models into the product conversation instead of leaving them as demo hardware toys.Read more
Backpropagation is bookkeeping with consequences
By: Jeremy London
Published: Apr 12, 2024 // Category: deep-learning-basics
Description: A slower explanation of gradients, loss, and weight updates for people who want the mechanism without the fog machine.Read more
An MLP is a stack of small calculations
By: Jeremy London
Published: Apr 4, 2024 // Category: deep-learning-basics
Description: A patient walkthrough of multi-layer perceptrons without pretending the acronym is more important than the arithmetic.Read more
Batch processing stops the lesson from lying
By: Jeremy London
Published: Mar 28, 2024 // Category: deep-learning-basics
Description: Why running several inputs at once gives a more honest picture of what a model is doing.Read more
OpenTelemetry profiling brought performance closer to the main stack
By: Jeremy London
Published: Mar 27, 2024 // Category: ai-operations
Description: Profiling moving into the OpenTelemetry conversation made performance feel closer to traces, metrics, and logs.Read more
Hidden layers are just intermediate guesses
By: Jeremy London
Published: Mar 21, 2024 // Category: deep-learning-basics
Description: A plain-language explanation of hidden layers, why they help, and why the word hidden makes them sound stranger than they are.Read more
Four neurons are enough to expose the wiring
By: Jeremy London
Published: Mar 14, 2024 // Category: deep-learning-basics
Description: A grounded walkthrough of a small neural network where the important part is the shape of the calculation.Read more
A single neuron is less mysterious when you run the numbers
By: Jeremy London
Published: Mar 7, 2024 // Category: deep-learning-basics
Description: A slower, more practical walkthrough of weights, bias, ReLU, and the first little model worth drawing by hand.Read more
Claude 3 gave the model picker real tradeoffs
By: Jeremy London
Published: Mar 5, 2024 // Category: model-watch
Description: Claude 3 turned Anthropic's lineup into a practical choice between speed, price, and capability instead of one vague assistant.Read more
Long context changed retrieval expectations
By: Jeremy London
Published: Feb 19, 2024 // Category: deep-learning-networks
Description: Gemini 1.5 made long-context work feel less theoretical, but it did not remove the need for retrieval judgment.Read more
Gemini 1.5 made long context the main event
By: Jeremy London
Published: Feb 16, 2024 // Category: model-watch
Description: Gemini 1.5 Pro changed the Google story from catch-up to a specific advantage: very large context across mixed media.Read more
Figma Dev Mode made components easier to inspect
By: Jeremy London
Published: Feb 7, 2024 // Category: design-systems
Description: Dev Mode made design-system drift easier to see because the design file became more inspectable by developers.Read more
Feature flags are product memory
By: Jeremy London
Published: Jan 17, 2024 // Category: product-engineering
Description: Flags are rollout switches, but they also record uncertainty, ownership, and the path back.Read more
Mixtral made open weights feel usefully weird
By: Jeremy London
Published: Dec 12, 2023 // Category: model-watch
Description: Mixtral 8x7B made sparse open-weight models feel like a practical path instead of a research curiosity.Read more
Astro 4 brought content sites closer to something bigger
By: Jeremy London
Published: Dec 12, 2023 // Category: web-engineering
Description: Astro 4 landed near the end of 2023, and it made static content work feel more like a serious application surface.Read more
Gemini 1 arrived with a multimodal reset
By: Jeremy London
Published: Dec 7, 2023 // Category: model-watch
Description: Google's Gemini 1.0 launch made the next model fight feel multimodal, tiered, and product-shaped.Read more
Offline-first is a product promise
By: Jeremy London
Published: Nov 8, 2023 // Category: local-first-software
Description: Offline capability is not a storage trick. It is a promise about what the user can trust when the network disappears.Read more
GPT-4 Turbo made context a product budget
By: Jeremy London
Published: Nov 7, 2023 // Category: model-watch
Description: OpenAI's DevDay launch made bigger context windows feel like something product teams had to price, test, and explain.Read more
Evals should have come before agent hype
By: Jeremy London
Published: Oct 30, 2023 // Category: deep-learning-principals
Description: Agent demos were already impressive in 2023, but the missing habit was measuring whether the system got better.Read more
Vector search needed product questions in 2023
By: Jeremy London
Published: Sep 26, 2023 // Category: advance-deep-learning
Description: The vector database rush made semantic search easy to demo and surprisingly easy to ship without product discipline.Read more
Design tokens need owners
By: Jeremy London
Published: Sep 14, 2023 // Category: design-systems
Description: Tokens are useful only when somebody owns the meaning, migration path, and exceptions.Read more
AI autocomplete raised the stakes for code review
By: Jeremy London
Published: Aug 18, 2023 // Category: developer-tools
Description: Autocomplete got good enough to change the texture of everyday programming, but review still has to own the final judgment.Read more
Figma Dev Mode changed handoff expectations
By: Jeremy London
Published: Jul 6, 2023 // Category: product-engineering
Description: Figma Dev Mode made handoff feel less like a screenshot exchange and more like an inspectable workflow.Read more
WebGPU made browser graphics feel serious
By: Jeremy London
Published: May 18, 2023 // Category: creative-coding
Description: Chrome 113 shipping WebGPU made high-performance browser graphics feel like a practical surface to learn from.Read more
Matrix multiplication is the first batch trick
By: Jeremy London
Published: Mar 4, 2023 // Category: deep-learning-basics
Description: A practical note on why neural network examples turn into matrices as soon as more than one example shows up.Read more