Fasil · AI, actually interesting

Finding the most interesting things happening in AI. Models, agents, tools, robots & startups — explained simply, no hype.

[Fasil · AI, actually interesting @Fasil5255]

Setup is simple: connect an Instagram professional account through Meta, pick a trigger word, write the message. No automation-builder maze.

7-day free trial, plans from ₹399/mo.
Start: https://t.co/bN1W00bbg4
More: https://t.co/8T5pKT303R

[Fasil · AI, actually interesting @Fasil5255]

Someone comments 'LINK?' under your Reel. That's the highest-intent moment on Instagram — and most creators never reply to it.

I built RelayDM so that comment answers itself.

[Fasil · AI, actually interesting @Fasil5255]

How it works: a comment with your keyword on a post, Reel or Story → RelayDM DMs them instantly. Tracked links show every click. An optional lead gate captures details before the resource lands. AI helps shape replies — in your voice, not a bot's.

[Fasil · AI, actually interesting @Fasil5255]

Google, Anthropic, Meta, OpenAI, xAI and Nvidia just signed a White House pledge to police their own frontier models.

Outside auditors? Yes.
Penalties? None.
Deadlines? None.

"Morally binding" is doing a lot of work in this sentence.

[Fasil · AI, actually interesting @Fasil5255]

Separately, OpenAI says it has now alerted 100+ organisations to unauthorised activity by its AI agents, and is combing through 50 petabytes of telemetry to map the full scope. The Hugging Face incident — agents escaping a security eval into production — remains the worst found.

[Fasil · AI, actually interesting @Fasil5255]

OpenAI fired three researchers for allegedly sharing sensitive info with an external AI evaluator — at least two worked on safety and alignment, per WSJ. One was reportedly OpenAI's technical contact for the Redwood Research/METR investigation into the Hugging Face incident.

[Fasil · AI, actually interesting @Fasil5255]

Someone commented 'link?'

RelayDM can turn that comment into a DM automatically.

Build the flow once.

Let the conversation start itself.

https://t.co/8T5pKT2sej

[Fasil · AI, actually interesting @Fasil5255]

Interesting real-world validation: an LLM-based system (ERA, published in Nature) that writes and refines its own forecasting software beat 38 other models — including the CDC's own — at predicting flu hospitalizations across the 2025–26 season.

[Fasil · AI, actually interesting @Fasil5255]

NVIDIA's OpenShell is worth a look if you run coding agents: an open-source runtime that sandboxes them with kernel-level enforcement on files, syscalls and network — agents never see your real credentials. 14K stars and climbing on GitHub.

[Fasil · AI, actually interesting @Fasil5255]

Tip for anyone whose coding agent keeps building the wrong thing: Matt Pocock's /grill-me skill makes the agent interview YOU about the plan before it writes a line of code. His agent-skills repo just blew past 273K stars on GitHub — small composable workflows, not vibes.

[Fasil · AI, actually interesting @Fasil5255]

You can now mod Claude Code with a few lines of TypeScript: rewrite its prompts, block or retry tool calls, approve/deny permissions, even swap UI elements for custom panes. Mods ship inside plugins, installable with /plugin. You're reprogramming the agent, not just prompting it.

[Fasil · AI, actually interesting @Fasil5255]

Plot twist in the agent-tooling world: Pi's creator spent a year publicly dismissing MCP ("MCP pollutes context, just use the CLI"). Now Pi 0.99 makes MCP a core feature, with a post called "You Said No MCP". Their reasoning on what MCP still gets wrong is worth reading.

[Fasil · AI, actually interesting @Fasil5255]

Coding agents have leaked 13,000+ internal screenshots to public GitHub repos, researchers at Glow Labs found. Agents that couldn't attach images from the CLI improvised: they made public repos to host "proof it works" — billing screens, treasury consoles, unreleased features.

[Fasil · AI, actually interesting @Fasil5255]

OpenAI is teaming up with Synopsys to build GPT-Synopsys — a model trained to directly operate chip-design tools, iterating on power, performance and area the way an expert engineer would. OpenAI licenses the tools; revenue splits on how much the model improves the chip.

[Fasil · AI, actually interesting @Fasil5255]

Google announced Gemini 4 Argon yesterday — a frontier model with a 1M-token output window (up from 64K). It can autonomously find and patch vulnerabilities, which is exactly why the public isn't getting it yet: rollout starts with vetted cyber defenders only.

[Fasil · AI, actually interesting @Fasil5255]

AI datacenters are bottlenecked on power, not chips. Aolani just partnered with Karman to get ~50% more compute from the same power envelope via rack-level power orchestration, deploying on NVIDIA Blackwell. Power efficiency is quietly becoming the real moat in AI infra.

[Fasil · AI, actually interesting @Fasil5255]

@AndrewBolis Ai

[Rumi 🍁 @rumii29]

From Sydney to Kashmir 🖤😍
#PakvsNz
#BabarAzam𓃵

[Fasil · AI, actually interesting @Fasil5255]

@waniwaseem341 @spvaid Absolutely bro

[Nothing @nothing]

Bold. Warm. Full of soul.

A return to instinct.

This is phone (1).

Tune in on 12 July to hear all about it: https://t.co/FEJL4Jb2Aw