Ex-Microsoft Engineer! ๐ค AI news, tools & tips summarized for you ๐ค From Claude to Cursor, plugins to trends ๐ ๏ธ Stay sharp, skip the noise ๐ฐ
London
- Tweets 7,205
- Following 705
- Followers 686
- Likes 247
Jony Ive is designing a hockey puck for OpenAI.
It costs $300, runs on a battery, and has no display. It arrives in 2027.
If the UX is just a better version of the voice assistants we already ignore, it is going to be a very expensive paperweight.
39
Stop switching apps to check live match scores.
Claudinho is a 2k-starred CLI tool that pipes real-time standings and data directly into your terminal. It works offline, needs no API keys, and puts the game status right where your code lives.
1
31
The brain drain at Google is now critical.
Jeff Dean and the core team are out. When you lose the people who built the infrastructure, you lose the ability to ship.
The AI arms race just got very lopsided.
19
Customer support stacks are about to get gutted.
Ringg is reporting 65% of customer calls resolved by AI agents. The real number is the cost: 90% cheaper than using GPT-4.1.
When the price drops that hard, "human-in-the-loop" isn't a feature anymore. It's just legacy overhead.
17
OpenAI just nuked Apple's lawsuit with an email dump.
The defense is brutal. Apple's lawyers apparently confused two people just because they share a last name. Receipts show Apple employees were actually the ones reaching out for help the whole time.
27
Reasoning is officially a commodity.
OpenAI just gave free users unlimited access to GPT-5.6 Luna with a built-in Think button. Meanwhile, Pro users get a granular thinking slider for Sol.
The era of gated reasoning is closing.
17
Stop clicking through CapCut to build your shorts.
capcut-cli lets you programmatically edit projects via JSON. It handles timeline assembly and subtitle generation locally, with no servers or API calls required.
1
16
36% fewer tokens.
That is the efficiency gap between Opus 5 and GPT-5.6 Sol for complex Excel modeling.
If you are running agentic finance workflows, your cost basis just shifted.
1
1
16
Itโs happening.
OpenAI throttled the Astra model after it demonstrated the ability to perform full end-to-end cyberattacks on hardened targets. Development is paused.
The line between a coding assistant and a security weapon just disappeared.
1
20
The breakup talk is over.
OpenAI just cemented GPT-5.6 as the engine for Microsoft Copilot 365.
If you were worried about your infra becoming obsolete, you can stop. The Microsoft-OpenAI stack is locked in for the next cycle.
1
23
Nvidia has a new rival.
AMD just launched the Helios rack-scale system. It is shipping today, with Microsoft, OpenAI, and Anthropic already deploying it at gigawatt scale.
The hardware market just became a two-horse race.
1
21
Voice-to-text always fails on technical terms until you add project-level context.
I switched to @github/vimalk78/dictate. It is a Linux daemon that supports per-project vocab hints and offloads transcription to your LAN GPU.
17
The "read-the-web" era of AI agents just hit a wall.
Amazon is officially blocking Meta's Muse agent. If your product relies on scraping public data to function, your uptime is now at the mercy of site owners.
The era of free access is closing.
18
Stop guessing which layers to prune in your local models.
Treat your LLM like a physical system instead. A new framework uses Ising optimization to mathematically solve which blocks to remove without breaking the model. Rigorous compression is finally here.
17
Meta just released Muse Glimmer.
It is a 30B parameter model built for local agents. Apache 2.0 license. It handles tool calling and code debugging entirely offline.
The era of renting private intelligence from the cloud is officially optional.
28
The industry is obsessed with LLM-only support agents.
Omilia just raised $67M to bet against it. They are scaling non-LLM automation, focusing on unit economics instead of just throwing models at the problem.
Sometimes you need a knife, not a bazooka.
27
Your terminal is a loaded gun.
One bad instruction and Claude Code can wipe your drive. CC Safety Net acts as a gatekeeper, intercepting destructive commands like `rm -rf` by analyzing intent. It kills the process before execution, even when buried in recursive wrappers.
1
34
We treat scientific software as ground truth. Anthropic just proved it is often just a guess. They found 240 structural errors in critical drug binding sites.
Anthropic ran a model against every deposited ribosome structure and found 240 that fold in ways no force field allows, with 11 flagged regions sitting at drug binding sites where published arguments rest on atoms placed by software rather than data.
23
Specialized medical LLMs are usually lobotomized. You train for clinical data, and reasoning tanks.
Cura 1T fixes this. It uses a human-gated self-evolution loop. It nails clinical reasoning and EHR tools without sacrificing the base model's intelligence.
9