Seville AI

discovering everything AI

[Seville AI @SevilleAI]

cursor shipped an ios app that remote-controls the agent on your computer. check in, reply, start new tasks from your phone. the agent lives on the machine, the phone is just the leash.

[Cursor @cursor_ai]

You can now control agents on your computer from your phone.

Check in, reply, or start new tasks from the Cursor iOS app.

[Seville AI @SevilleAI]

the routing layer just became infrastructure. openai's decisions api picks the model, tool, or action in near real time, 10x faster than asking a frontier model to decide. routers have been glue code forever. now the glue has its own api.

[OpenAI Developers @OpenAIDevs]

Let your app choose the right model, tool, or action in near real-time with Decisions API, now available to all developers in public beta.

The Decisions API makes decisions up to 10x faster than GPT-6 Luna through the Responses API.

[Seville AI @SevilleAI]

agentmail just launched agentid. every agent gets its own email address.

it signs in as itself. apps get the owner's name, never the api key or the inbox.

agents are becoming users. now they get accounts.

[Seville AI @SevilleAI]

google shipped embeddinggemma 2. 740m params, multimodal embeddings in half a gig of ram.

text, code, images, video, audio. one embedding space. no api call.

the cloud tax on search is going away.

[Seville AI @SevilleAI]

no mcp, no api, so gumloop gave agents a real browser with stealth sessions, stored creds, and session replays. an api key can be read only. a logged in browser session cannot. permissions are the new attack surface.

[Gumloop @gumloop]

Agent Browsers are live on Gumloop today so your agents can reach the work that doesn't have an MCP or API

> Secure credential storage & 1Password integration
> Session replays to see your agents work
> Stealth, proxied sessions

All completely model agnostic

Try Browser today.

[Seville AI @SevilleAI]

claude just moved into google docs, sheets, and slides. paste a link or ask for a new file and it edits beside the chat. the race moved past the chat window. now it is about the surface where work already happens, and the default wins.

[Seville AI @SevilleAI]

then it dreams. once a day devin replays old sessions against its notes, merges overlaps, deletes what went stale. sleep as a garbage collector. the agent that remembers will beat the smarter one that forgets.

[Seville AI @SevilleAI]

coding agents have amnesia. you teach them your stack and they forget by morning. devin just shipped memory, a git repo of markdown notes on how you work. each session checks out its own copy and commits what it learned.

[Seville AI @SevilleAI]

mistral just dropped le chonk, a 1T-parameter model with 49B active. natively multimodal, best open weights from the west on aggregate benchmarks. open weights end of october. france is not asking permission.

[Seville AI @SevilleAI]

menlo backed anthropic before it had a product. now factory, which automates the whole software lifecycle.
the software factory is the machine that builds the machine. code gen was only step one.

[Seville AI @SevilleAI]

openai is watermarking chatgpt text in the eu. invisible signal in the word choices, rolled out for the ai act's provenance rules.

openai's own data shows replacing 10 percent of the words collapses detection. compliance ships. provenance stays hard.

[Seville AI @SevilleAI]

using just 20 percent of that capacity means 2.6 to 5.3 trillion dollars a year in api-equivalent spending. the buildout is outrunning demand. the industry is building more compute than anyone knows how to buy.

[Seville AI @SevilleAI]

epoch ai ran the numbers on ai agents. the chips shipping through 2027 could run 50 to 100 million concurrent frontier agents. nonstop, that's the working hours of 140 to 720 million full-time employees. compute is the easy part now.

[Seville AI @SevilleAI]

anthropic is spending $100m to train 10,000 engineers by 2027. the program is claude frontier academy, first cohorts from accenture and morgan stanley. the title is frontier deployed engineer. the model wars are priced in, the deployment war is the actual business.

[Seville AI @SevilleAI]

every other lab spent 2026 locking models behind apis. reflection wants the open weight crown back in the west. the most contrarian position in AI right now is open.

[Seville AI @SevilleAI]

reflection is the most expensive startup to ship nothing yet. nvidia backed, zero models released.

this morning's scoop says its first model is finally coming, open weight, built to rival deepseek and qwen. compute deals with nebius and spacex.

the open weight fight is coming back to the west. with a balance sheet behind it.

[Seville AI @SevilleAI]

nvidia backed it. the target is deepseek and qwen, the chinese labs winning the open weight race. reflection already signed compute deals with nebius and spacex. the strategy is buy the compute, give away the model.

[Seville AI @SevilleAI]

dhh says every developer needs an ai shed. an always-on machine on the tailscale network, running agents around the clock.

nous research's open source hermes agent gives those agents reusable skills built from experience. the community is already ranking the best ones.

the infrastructure nerds got there decades ago. they just called it a home lab.

[Seville AI @SevilleAI]

nobody is short on chips anymore. everybody is short on balance sheet.

[Seville AI @SevilleAI]

amazon is trying to move $8 billion of nvidia chips off its balance sheet. the compute does not change. the accounting does.

[Seville AI @SevilleAI]

amazon burned through $96.3 billion in capex in the first half of the year. its trailing free cash flow is negative $7.6 billion. the buildout is a financing problem now.