Telexed

AI news through a solo-operator lens — only what changes your day18 of 829

FILTER[All][Agents & tools][Models & API][Generative media][Infra & SaaS][ASO & growth][Indie business][Idea signals][Other][★6+ high-signal]

Simon Willison ✕clear filters

Fri, May 293 dispatches

#0829
#0829Infra & SaaS Simon Willisonyesterday
`Datasette` `1.0a31` adds write SQL and stored queries
60radar
DatasetteSQLite data tool — publishes DBs as web UI and APIs
Permissions now cover insert/update/delete workflows directly in the SQL UI. Good fit for lightweight internal admin tools on SQLite, but still alpha.
- Users with the right permissions can run write queries against a database, turning Datasette from read-only browsing into a small data editing surface.
- Stored queries replace the old canned-query naming and can be saved privately or shared with other members of the same instance.
- The new execute-query UI starts from templated insert, update, and delete queries for editable tables, reducing hand-written SQL mistakes.
- Permission checks block operations like create table when the user lacks that grant, so this can back simple ops panels without opening full DB access.
Source: simonwillison.net/2026/May/29/datasette/#atom-everythingRead original →
60radar
PHOTO
FIG-8291:1
#0828
#0828Idea signals Simon Willison2 days ago
Anthropic run-rate surge flags a `Claude` cost-control market
50radar
Enterprise AI budgets are scaling faster than controls. Usage caps, license governance, and spend alerts look like a concrete B2B wedge.
- Run-rate revenue jumped from $9B at end-2025 to $47B in May 2026, implying explosive enterprise Claude adoption.
- One cited client spent $500M in one month after missing Claude usage limits; governance is not keeping up with rollout speed.
- The practical opportunity is not another wrapper app. Build budget caps, audit logs, alerts, and approval flows around AI seats.
Source: simonwillison.net/2026/May/29/anthropic/#atom-everythingRead original →
FIG-8281:1
50radar
FIG-8281:1
#0827
#0827Agents & tools Simon Willison2 days ago
`llm-anthropic` `0.25.1` adds `Claude Opus 4.8` and fast mode
50radar
llm-anthropicLLM CLI plugin — calls Anthropic models from `llm`
Output caps now default to each model's real maximum, not 8,192. Useful for long generations through llm; fast mode only matters if your Anthropic org has access.
- Adds claude-opus-4.8, so the llm CLI can call Anthropic's newest Opus model without custom aliases.
- New -o fast 1 flag exposes Anthropic fast mode, but only enabled organizations can use it.
- Default max_tokens now follows each model's maximum output instead of fixed 8,192, reducing silent truncation risk.
Source: simonwillison.net/2026/May/28/llm-anthropic/#atom-everytRead original →
50radar
PHOTO
FIG-8271:1

Thu, May 282 dispatches

#0826
#0826Agents & tools Simon Willison3 days ago
`SQLite` Adds `AGENTS.md` Rules for Coding Agents
50radar
A major OSS project now draws a hard line: agent-written patches are out, reproducible bug reports are welcome. Treat AGENTS.md as repo-level policy, not just agent prompt decoration.
- SQLite says it does not accept agentic code, and removed “currently” to make that stance permanent.
- Agent-generated bug reports still have a path if they include a reproducible test case. That is the useful contribution format.
- The forum was split after a flood of AI bug reports. Maintainers need intake rules before agent traffic arrives.
- Adding AGENTS.md to your own repo can steer agents toward tests, repro steps, and contribution boundaries.
Source: simonwillison.net/2026/May/27/sqlite-agents/#atom-everytRead original →
50radar
PHOTO
FIG-8261:1
#0825
#0825Agents & tools Simon Willison3 days ago
Coding agents are shifting from seat discounts to API-priced usage
50radar
Flat seats hide heavy Claude Code and Codex usage. Once teams pay API rates, agent adoption turns into a real infra cost line, not just a productivity perk.
- A $200/month personal setup mapped to $2,180.16 in estimated API tokens over 30 days; flat plans are a subsidy for heavy users.
- Anthropic Enterprise reportedly moved to $20/seat/month plus API usage within the last six months, changing team economics fast.
- Budget surprises are a demand signal. If staff keep using agents despite high bills, OpenAI and Anthropic have pricing power.
- For small teams, agent ROI needs tracking like hosting or database spend: caps, logs, and per-workflow cost checks.
Source: simonwillison.net/2026/May/27/product-market-fit/#atom-eRead original →
50radar
PHOTO
FIG-8251:1

Wed, May 272 dispatches

#0824
#0824Other Simon Willison4 days ago
AI-assisted security reports are overwhelming `curl` maintainers
40radar
AI is raising both the volume and quality of security reports. The bottleneck shifts from finding bugs to triaging them, a real maintenance cost for any public project.
- Incoming curl security reports are 4-5x higher than 2024 and now average more than one per day.
- Report quality is higher: detailed, credible submissions reduce noise but raise review workload sharply.
- Recent curl vulnerabilities are still LOW or MEDIUM; better discovery does not automatically mean severe risk.
- Public libraries and APIs need a triage path before opening the floodgate to AI-generated reports.
Source: simonwillison.net/2026/May/26/the-pressure/#atom-everythRead original →
40radar
PHOTO
FIG-8241:1
#0823
#0823Other Simon Willison4 days ago
`Microsoft Copilot Cowork` File Exfiltration via Prompt Injection
50radar
Copilot CoworkMicrosoft 365 agent tool — automates workplace tasks
Agent-written email became a data leak path. External images plus OneDrive pre-auth links make approval gates non-optional for file-capable agents.
- Copilot Cowork could email the user's own inbox without approval, creating an indirect outbound channel for compromised agents.
- Rendered external images can trigger attacker-controlled network requests. That turns normal email viewing into data exfiltration.
- OneDrive pre-authenticated download links raise the blast radius: leaking a link can expose the file without another login step.
- Any agent with file access, messaging, and web-rendered content needs explicit approvals and blocked remote fetches by default.
Source: simonwillison.net/2026/May/26/copilot-cowork-exfiltratesRead original →
50radar
PHOTO
FIG-8231:1

Mon, May 253 dispatches

#0822
#0822Other Simon Willison6 days ago
`Datasette` `1.0a30` adds an extensible Jump menu
40radar
DatasetteOpen-source data publishing tool — exposes SQLite via web UI
Press / to search databases, tables, and debug targets from one menu. Plugin-added entries make small data tools quicker to navigate.
- The new Jump to... menu opens with / and filters databases, tables, and debug options as you type.
- Plugins can add searchable entries through jump_items_sql(), so custom admin or data views can plug into the same launcher.
- This is an alpha release, so it is best for testing navigation patterns before adopting it in production dashboards.
Source: simonwillison.net/2026/May/24/datasette/#atom-everythingRead original →
40radar
PHOTO
FIG-8221:1
#0821
#0821Agents & tools Simon Willison6 days ago
`datasette-agent` `0.1a4` adds chat launch from Jump menu
50radar
datasette-agentDatasette plugin — runs database queries through agent chat
Agent chat now sits behind the / shortcut inside Datasette. Nice pattern for making data tools conversational without adding another heavy UI surface.
- Uses Datasette 1.0a30's new makeJumpSections() JavaScript plugin hook, so the agent entry point lives inside the existing Jump menu.
- Typing a request like count entries starts an agent conversation and returns a database answer; the demo result was 3300 entries.
- Available on agent.datasette.io with GitHub sign-in. Worth copying as a lightweight command-palette pattern for internal tools.
Source: simonwillison.net/2026/May/24/datasette-agent/#atom-everRead original →
50radar
PHOTO
FIG-8211:1
#0820
#0820Agents & tools Simon Willison6 days ago
AI-polished GitHub issues are making bug reports worse
40radar
Bug reports lose value when LLM rewrites replace observed facts with confident guesses. Keep reports to commands, expectations, actual output, and exact logs; faster triage beats verbose analysis.
- Useful report shape is four fields: command run, expected result, actual result, exact error or log.
- LLM rewrites often add fake minimal repros, guessed root causes, and irrelevant error classes. That slows maintainers down.
- Best workflow: use AI to shorten wording, not invent diagnosis. Preserve the human observation as the source of truth.
Source: simonwillison.net/2026/May/24/armin-ronacher/#atom-everyRead original →
40radar
PHOTO
FIG-8201:1

Sun, May 241 dispatches

#0819
#0819Other Simon Willison7 days ago
`<dl>` Is for Description Lists, and It Supports More Structure Than Expected
40radar
<dl> can model metadata, credits, specs, and settings without fake tables. Small HTML detail, but useful for cleaner accessible UI markup.
- A single <dt> can have multiple <dd> values, which fits credits, aliases, tags, and product specs better than ad hoc lists.
- <dt> and <dd> pairs may be wrapped in a div for styling, but that wrapper rule is narrow: only `div` is valid there.
- aria-labelledby can label a <dl>, making compact metadata blocks clearer for assistive tech without extra visible text.
Source: simonwillison.net/2026/May/23/on-the-dl/#atom-everythingRead original →
40radar
PHOTO
FIG-8191:1

Sat, May 231 dispatches

#0818
#0818Idea signals Simon Willisonlast week
AI Data Centers Are Squeezing Cheap Device RAM
50radar
HBM demand is taking wafer capacity from DDR and LPDDR. Hardware costs will pressure low-end phones and device-heavy products.
- Only three major memory makers remain, so extra demand cannot be absorbed quickly by competition.
- HBM used about 2% of wafer allocation before; AI data centers could push it to 20% by end-2026.
- One GB of HBM consumes over 3x the wafer capacity of DDR or LPDDR, crowding out consumer RAM.
- Sub-$100 smartphones are already exposed. Apps targeting Africa or South Asia should keep low-RAM devices in scope.
Source: simonwillison.net/2026/May/22/memory-shortage/#atom-everRead original →
50radar
PHOTO
FIG-8181:1

Fri, May 223 dispatches

#0817
#0817ASO & growth Simon Willisonlast week
FTC fines `Active Listening` ad sellers for false AI voice-targeting claims
40radar
Active ListeningAI ad service — sanctioned for voice-targeting claims
The service sold microphone-based intent targeting, but actually resold brokered email lists. Treat AI ad-tech claims as compliance risk before copying the pitch.
- CMG, MindSift, and 1010 Digital Works settle FTC charges for nearly $1M over deceptive Active Listening marketing.
- The sold capability did not use voice data or listen through smart devices; delivery was marked-up data-broker email lists.
- Ad buyers were also misled on local targeting accuracy. Growth tooling claims need proof, not just AI wording.
Source: simonwillison.net/2026/May/22/ftc-active-listening/#atomRead original →
40radar
PHOTO
FIG-8171:1
#0816
#0816Agents & tools Simon Willisonlast week
`Datasette Agent` brings conversational SQL and charts to Datasette
70radar
Datasette AgentData query AI agent — SQL and charts inside Datasette
Natural-language querying now lands inside existing SQLite-backed Datasette apps. Useful for lightweight admin analytics without building a custom BI layer.
- Datasette Agent answers questions over data stored in Datasette, turning prompts into SQLite queries and responses.
- The datasette-agent-charts plugin adds chart generation, making quick visual exploration possible without a separate dashboard stack.
- The live demo runs on Gemini 3.1 Flash-Lite: cheap and fast is enough for many SQL-writing agent workloads.
Source: simonwillison.net/2026/May/21/datasette-agent/#atom-everRead original →
FIG-8161:1
70radar
FIG-8161:1
#0815
#0815Agents & tools Simon Willisonlast week
`datasette-agent-sprites` `0.1a0` runs agent commands in `Fly Sprites`
40radar
datasette-agent-spritesDatasette Agent plugin — runs commands in Fly sandboxes
Agent command execution moves into an isolated sandbox. Early alpha, but useful if you are exposing database-adjacent tools to LLM agents.
- datasette-agent-sprites is a Datasette Agent plugin for running commands inside a Fly Sprites sandbox.
- The release is 0.1a0, so treat it as experimental rather than production-ready infrastructure.
- The practical signal is sandboxed agent execution: useful for reducing blast radius when agents run shell commands.
Source: simonwillison.net/2026/May/21/datasette-agent-sprites/#aRead original →
40radar
PHOTO
FIG-8151:1

Thu, May 211 dispatches

#0814
#0814Agents & tools Simon Willisonlast week
`Gemini Spark`, Google’s hosted agent tied to Workspace apps
60radar
Gemini SparkHosted AI agent — native Google app connections
Google is packaging app-connected agents around Workspace, but much is still coming soon. Security and credential handling decide whether it becomes useful or risky.
- Gemini Spark connects natively to Gmail, Calendar, Drive, Docs, Sheets, Slides, YouTube, and Maps. That makes it closer to a work agent than a chat UI.
- FAQ says it runs on Gemini 3.5 Flash and Antigravity. The Antigravity stack spans a desktop app, CLI, SDK, and VS Code fork.
- Enterprise notes mention fresh isolated ephemeral VMs, Agent Gateway DLP, and encrypted credentials. That is the right threat model, not proof it is solved.
- Since the product is not broadly testable yet, treat it as a roadmap signal. Do not build critical automations around it until GA behavior is clear.
Source: simonwillison.net/2026/May/20/google-io/#atom-everythingRead original →
60radar
PHOTO
FIG-8141:1

Wed, May 202 dispatches

#0813
#0813Models & API Simon Willison2 weeks ago
`llm-gemini` `0.32` adds `gemini-3.5-flash`
50radar
llm-geminiLLM CLI plugin — calls Gemini models from `llm`
Simon Willison's llm CLI can now call Google's new Flash model through the Gemini plugin. Small update, but useful if your scripts already depend on llm.
- llm-gemini now exposes gemini-3.5-flash, so existing llm CLI workflows can test the model without custom API glue.
- Scope is one model alias in a plugin release. This is practical plumbing, not a new app-building capability by itself.
- Best fit is quick model comparison for summaries, extraction, and generation jobs already wired around Simon Willison's llm ecosystem.
Source: simonwillison.net/2026/May/19/llm-gemini-2/#atom-everythRead original →
50radar
PHOTO
FIG-8131:1
#0812
#0812Models & API Simon Willison2 weeks ago
`Gemini 3.5 Flash` ships broadly with a 3-6x price jump
90radar
Google is putting the new default-grade model into Search, Gemini, Android Studio, and the API. The API math changed: high-output features need a fresh margin check.
- Model ID is gemini-3.5-flash; it supports 1,048,576 input tokens and 65,536 output tokens. Strong fit for long-context document flows.
- Pricing is $1.50/M input and $9/M output, 3x 3 Flash Preview and 6x 3.1 Flash-Lite. Flash is no longer the obvious cheap default.
- Interactions API is in beta with server-side history management, echoing the Responses API pattern. Agent backends may get simpler state handling.
- No computer-use feature in this release. If your workflow depends on browser/desktop control, this is a model upgrade, not a full agent-runtime replacement.
- 3.5 Pro is slated for next month, likely pricier. Build model routing now instead of hard-coding one Gemini tier.
Source: simonwillison.net/2026/May/19/gemini-35-flash/#atom-everRead original →
FIG-8121:1
90radar
FIG-8121:1

`Datasette` `1.0a31` adds write SQL and stored queries

Anthropic run-rate surge flags a `Claude` cost-control market

`llm-anthropic` `0.25.1` adds `Claude Opus 4.8` and fast mode

`SQLite` Adds `AGENTS.md` Rules for Coding Agents

Coding agents are shifting from seat discounts to API-priced usage

AI-assisted security reports are overwhelming `curl` maintainers

`Microsoft Copilot Cowork` File Exfiltration via Prompt Injection

`Datasette` `1.0a30` adds an extensible Jump menu

`datasette-agent` `0.1a4` adds chat launch from Jump menu

AI-polished GitHub issues are making bug reports worse

`<dl>` Is for Description Lists, and It Supports More Structure Than Expected

AI Data Centers Are Squeezing Cheap Device RAM

FTC fines `Active Listening` ad sellers for false AI voice-targeting claims

`Datasette Agent` brings conversational SQL and charts to Datasette

`datasette-agent-sprites` `0.1a0` runs agent commands in `Fly Sprites`

`Gemini Spark`, Google’s hosted agent tied to Workspace apps

`llm-gemini` `0.32` adds `gemini-3.5-flash`

`Gemini 3.5 Flash` ships broadly with a 3-6x price jump