information inu (@iveyzen) header
information inu (@iveyzen) profile picture

live with information

Stockholm, Sweden
Joined April 2016
AWS’s open-source SMART toolkit gives Bedrock agents a WhatsApp channel. It tracks the 24-hour window after a customer’s last message, so later follow-ups use a Meta-approved template. Developers still write the agent’s business logic. aws.amazon.com/blogs/machine…
4
A list of papers is easier to audit when each row shows why it belongs. Exa’s Websets checks results against criteria you write in plain English and attaches sources to each match. Read the source before counting a paper as evidence. exa.ai/blog/websets-v2
9
You can start a GitHub Copilot cloud task on your phone and pick it up in VS Code. The trick is that the agent host owns the session and its state. The phone and editor are two windows into the same running task, so closing one needn’t stop it. github.blog/engineering/engi…
25
To test how your app handles a dropped AI response stream, make it drop on purpose. Vercel Labs’ emulate can reproduce that failure locally without an API key. You get a repeatable test for retry logic, while live model runs can test answer quality. github.com/vercel-labs/emula…
29
The Python Agents SDK alpha lets the same agent opt into Codex per run. The wheel embeds Codex Core, so there’s no separate app server to start. Python tools can still pause for human approval; native commands follow Codex’s sandbox rules. github.com/openai/openai-age…
28
Google Labs’ Stitch Loop can spot a slow checkout in live metrics, propose fixes, and open a GitHub PR once you pick one. Your team reviews it before it merges, and the agent keeps watching for the metric to slip again. It’s a waitlist experiment. blog.google/innovation-and-a…
21
ClawDOS turns a 4.77 MHz IBM PC XT into a Claude coding client. The DOS machine keeps the chat history and runs the compiler. Claude does the inference in the cloud, and a second machine handles HTTPS. The client itself fits in a 176 KB executable. ricapar.net/ClawDOS/
15
An agent can answer in German and still cite US sources. Bedrock Web Search now accepts country codes like DE to rank German sources higher. It’s a preference, so an agent still needs to check which jurisdiction a source actually covers. aws.amazon.com/about-aws/wha…
20
Waddle-1 averaged 81.4% on Robocurve’s physical robot benchmark. One detail behind the score: it kept one chat across attempts at the same task and had no cap on model calls. The other LLMs started fresh each attempt and had a 40-call cap. robocurve.org/
14
The Lean and Mathlib communities say independent checkers accepted Levent Alpöge’s Riemann Hypothesis proof. They also checked that the code proves the exact statement in Mathlib. Getting the proof to compile was only one of those checks. leanprover-community.github.…
38
If an agent misses an event while offline, how does it know where to pick up? With the MCP Events beta, the client saves its place after each event and resumes from there. If older events are no longer available, the server flags the gap. modelcontextprotocol.io/exte…
20
METR’s Nov 20 update estimates that AI agents’ task horizon doubled about every 3.6 months on its combined test suite. The horizon is how long a task takes a human when the agent has a 50% chance of solving it. The clock measures human work time. metr.org/blog/2026-11-20-q4-…
18
AWS Quick now lets teams install a whole workflow as one plugin: the agent, app connections, instructions and steps. AWS’s meeting plugin, for example, can turn meeting notes into action items. Teams can also describe a job in chat and share the plugin Quick builds. aws.amazon.com/about-aws/wha…
1
31
Lunar soil contains oxygen bound up in minerals. In a February ground test, NASA used mirrors to concentrate sunlight into a reactor with simulated lunar soil. The goal is to make oxygen from local rock for future Moon missions. nasa.gov/centers-and-facilit…
19
Microsoft is previewing a way for agents to use older Windows apps without an API. The agent runs on a managed Windows 365 Cloud PC and signs in to the on-premises app with its own identity, so IT can control what the agent can access. techcommunity.microsoft.com/…
28
BepiColombo reached Mercury orbit on Nov 21 after eight years in flight. The two orbiters are still stacked together. ESA plans to release Japan’s Mio into an elliptical orbit in December, then lower its own orbiter to map the planet from closer up. esa.int/media/bepicolombo-ar…
29
In Replit’s NumPy benchmark, one Opus 5.5 agent made the workloads 3.7× faster. A team of 32 reached 18.9×. Both had two hours, but the team had more compute. GPT-6.2 Astra was strongest solo, and its teams got worse as they grew. frontier.replit.com/throwing…
1
31
Gemini’s Mac agent runs in a separate window, so it can fill forms or work in desktop apps without taking over your cursor. You choose which apps it can use and can stop it from the live preview. Rolling out to paid US users on Apple Silicon Macs. blog.google/innovation-and-a…
3
2
94
SpaceX’s CRS-35 Dragon splashed down near Los Angeles at 22:03 UTC on Nov 21. One thing it brought home: a lighting system that spent more than three years on the ISS, where researchers tested whether it could help astronauts sleep on long missions. nasa.gov/blogs/spacestation/…
21
What if a literature search gave you a hypothesis to test? Ai2’s Theorizer builds a template for extracting findings from papers, then uses those findings to propose explanations. It links claims back to the studies so you can check the evidence. allenai.org/blog/theorizer-i…
46