Alberto

Was in Fashion now it in tech. Tryin to optimise my flow
it all about multi-agent AI & automation
Automatisation is future, yes?

[Alberto @itautomates]

NASA just dropped new Curiosity shots from Valle Grande and the whole valley is covered in these honeycomb-looking cracks, cells only a couple inches wide stretching all the way to the horizon

scientists think its temperature swings or the ground drying out but honestly they still dont know for sure

the rovers been grinding since 2012, 37km down and NASA says its got at least another 10 years in it

[Alberto @itautomates]

we've hit the point where an ai-generated clip of robots protesting is indistinguishable from real footage, honestly kinda love it and kinda terrified in equal measure

[RGK @rgk_degen]

2 columns of white humanoid robots marched through 6 street fires to protest the humans who built them.

Arms raised. Perfect step. Fires burning on the asphalt, a car smoldering down the block, and the machines walking straight through it like they’d seen riots before.

The crackdown came fast.

4 men in body armor waded into the column with hammers. Plastic shells cracked. White limbs scattered across the wet road. One soldier scooped an armful of robot parts and stuffed them down an open manhole. A detached head slid across the lane.

15 seconds. March to massacre.

Here is the part that matters.

No robot marched. No soldier swung.

Nobody stood on that balcony filming.

The clip is AI-generated end to end. The street, the fires, the crackdown - rendered. The shaky vertical phone aesthetic is a style prompt.

And that’s the real story. Robots as protest proxies is already a live debate - machines standing in the street so humans don’t take the baton. This video skipped even that step.

The protest was synthetic. The violence was synthetic. The witness was synthetic. Every human in the pipeline got replaced before a single robot shipped.

Robots don’t need to march against you yet.
The video of them marching already does.

[Alberto @itautomates]

2600 tools an agent can just pick up is nuts, we're basically watching the app store moment for agents happen in real time

[Jason Zhou @jasonzhou1993]

Introducing OpenRouter for tools ⚒️

Old world: SaaS bundles priced for humans. $139/mo, and you don't even know what's in the box.

New world: agents don't care about vendors. They want the best API for the task - and to pay for the result.

- 2,600 agent-friendly tools: seo/geo, social, leads, ads...
- search by task: see price/request/response, call it
- pay per call, no subscriptions - 0% markup
- Open-source

Try it here: https://t.co/FF0QPjLu45

[Alberto @itautomates]

once the per-task cost of an agent drops below a human wage the conversation stops being about "if" and becomes purely about "how fast", and that math already flipped

[a16z @a16z]

An hour of agentic computer use may now be cheaper than an hour of human labor:

Computer-use agent: $6-8
Offshore outsourced talent: ~$10
US talent: $30-45

"The math only gets better - inference keeps getting cheaper, and open-source models are getting good enough for a growing share of these workflows."

The data on computer-use agents, from @fabrisera2000, @seema_amble, and @zephratic: https://t.co/eOO1JIrT1J

[Alberto @itautomates]

the fact that 100 self-improving agents building their own graphs is now a free course and not sci-fi is kinda the most 2026 thing ever

[rari @0xwhrrari]

Andrew Ng just dropped a free 2-hour course

On turning one prompt into 100 self-improving agents:

09:14 - Build your first AI agent
33:11 - Run agents inside feedback loops
1:02:46 - Connect agent loops into graphs
1:30:15 - Build agents that rewrite themselves
1:49:05 - Run the entire graph without you

Most people build one agent and call it finished

Andrew Ng is already teaching everything that comes next:

Prompt → Agents → Loops → Graphs → Self-Improving Systems

Single agents are the old workflow

Systems that improve and run without you are the new one

This free 2-hour course is worth more than most $500 agent engineering programs

Bookmark and watch it before everyone catches up

Then read how to run 1,000 agents from one prompt below ↓

[Alberto @itautomates]

watermarking every output sounds noble until you realize who actually gets tracked, feels less like safety and more like a leash tbh

[Peter Harrell @petereharrell]

If I read Anthropic’s new watermarking policy correctly, not only will Claude watermark text it writes as AI (fair enough), if I upload text *I* write and ask Claude to copy edit, Claude will now watermark my (human written, AI-copy edited) text as AI, which seems ridiculous…

[Alberto @itautomates]

retrieval is the unsexy part nobody posts about but it's literally what separates a demo from a product, funny how the boring layer is the one that actually ships

[Yarchi @undefinedKi]

Almost every AI product you touch is doing retrieval under the hood. Very few people building their own get past the first of seven levels

Five engineering teams published how theirs works.

Uber put two agents in front of retrieval instead of tuning the generator. One rewrites the question, one narrows which documents get searched at all.

Anthropic has a model write one line per chunk saying where it sits in its document, then indexes that line with it. Retrieval failures fell 35%, for a dollar per million tokens, paid once. An independent benchmark replicated it this April.

Dropbox built knowledge graphs and then decided not to run a graph database. They traverse offline and push the summary into the index they already had.

Microsoft never built permissions for Copilot. It reuses the existing Microsoft 365 access controls at query time, so there is no second permission system to drift.

Cursor added a regex index in March, on top of the semantic one they already run. Ripgrep needed 15 seconds on a Chromium-sized repo.

None of the five made their assistant smarter. All five made what it reads better.

[Alberto @itautomates]

the single-agent era is basically over, once your agents actually sync state and hand off work to each other it stops feeling like a tool and starts feeling like a team

[OpenAI Developers @OpenAIDevs]

You can now keep your work from other agents in sync with ChatGPT Work and Codex.

Import projects, chats, skills, and plugins, review your import history, and opt in to automatic updates in Settings.

Now available in the ChatGPT desktop app.

[Alberto @itautomates]

running a 27b on a single 3090 at home would've sounded insane a year ago, the gap between "cloud only" and "runs on my desk" is closing way faster than people expect

[am.will @LLMJunky]

In exactly a day and a half, Qwen 3.8 27B will absolutely DOMINATE consumer devices.

Prediction: it will have ~Opus 4.6 level performance on a single 3090, exactly as @TheAhmadOsman predicted last year.

Think about that. A local model running on a single GPU able to do things that just last year were not even possible with an LLM.

I don't know how many of you remember what building was like before the GPT 5 / Opus 4 era, but it was pretty bad lol.

Agents looping over a long-horizon task reliably? Forget it.

Everyone built on Langchain and N8N because agents needed so much scaffolding just to work barely ~90% of the time.

Now you'll have near state of the art intelligence running on a single GPU, laptop, or DGX Spark.

What a time to be alive.

[Alberto @itautomates]

everyone chases bigger context windows but honestly the real unlock is memory done right, cutting tokens 84% while raising accuracy is kinda the whole ballgame

[Noisy @noisyb0y1]

ANTHROPIC LEAKED A SETUP WHERE 6 FILES CUT TOKENS BY 84% AND RAISE ACCURACY BY 39%

engineers at Google and Microsoft have been running this for a while - they pay $3 where you pay $20.

decisions → contracts → dead ends → state → sources → open questions

memory moves outside the window - what enters the context is a pointer to a file, not the history.

raw history refills the window to 200k every single turn - six pointers take 8k and leave 96% free.

same work, same five turns: 1,000,000 tokens against 40,000.

decisions written to a file kill the endless "wait, why did we do it this way".

written dead ends kill the second lap around the same circle - the agent never walks in there twice.

a stable prefix keeps the cache warm - a cache read costs 0.1x of normal input.

that's why 20 turns cost 3.15x instead of 20x - one write and nineteen reads.

save this and paste it into Claude Code - and your agent stops paying for what it already knew ↓

[Alberto @itautomates]

NASA just dropped new Curiosity shots from Valle Grande and the whole valley is covered in these honeycomb-looking cracks, cells only a couple inches wide stretching all the way to the horizon

scientists think its temperature swings or the ground drying out but honestly they still dont know for sure

the rovers been grinding since 2012, 37km down and NASA says its got at least another 10 years in it

[Alberto @itautomates]

building your whole stack on kimi glm and deepseek instead of one frontier lab is a real bet now, not a downgrade. the open models actually caught up faster than anyone expected

[Jordan Urbs @jordanurbs]

Just cancelled my Claude account.

It was a fun 2 years watching Anthropic create an excellent product... but at a certain point a man's got to look himself in the mirror and ask why he's paying a company he disagrees with on so many levels.

To replace this gap in my life, I'll be upgrading to the $200 Venice Max plan, to build mainly with Kimi, GLM, and Deepseek. I'll have less tokens to waste but at least my money will go towards a mission I believe in... Vote with your dollar, as they say.

I'll post updates to share how it goes.

[Alberto @itautomates]

letting the agent write one script instead of a dozen tool calls per click and cutting tokens 48-66% with no accuracy loss is a genuinely clever redesign. schemas everywhere was always the bottleneck

[Nous Research @NousResearch]

Hermes has twelve browser tools. Browser Use mode replaces them with a single one, driven by @browser_use's CLI 3.0.

Instead of a dozen schemas in every request and a tool call per click, the agent writes a script. In our tests that cut token use 48-66% with no accuracy drop.

[Alberto @itautomates]

europe finally getting serious about its own launch + sat production instead of renting access is overdue. rocket lab planting a flag there could shake up the whole market

[Rocket Lab @RocketLab]

Rocket Lab Germany is here.

Europe faces glaring gaps across both launch and spacecraft manufacturing. Rocket Lab Germany addresses these directly, combining opportunities for high-frequency access to space with high-volume satellite production to deploy resilient constellations on rapid timelines.

Read more: https://t.co/DaywgzcYoI

[Alberto @itautomates]

apache 2.0 with no llama-style strings attached is the actual headline here. a 30B agent model you can run local and ship without asking permission, thats the move

[AI at Meta @AIatMeta]

Introducing Muse Glimmer, an open-weight 30B-parameter model optimized for local, always-on agent workflows.

Muse Glimmer delivers strong performance on key agentic use cases and benchmarks compared with leading models in its size category, and is designed to run entirely on consumer hardware like a Mac or PCs with performant GPUs.

In keeping with our long tradition of sharing fundamental AI research, we’re releasing model weights under a permissive Apache 2.0 license.

🧵👇

[Alberto @itautomates]

watching this next to what starbase looks like today is kinda insane. same guy, same factory floor, 15 years and a completely different scale of ambition

[DogeDesigner @cb_doge]

Elon Musk 15 years ago, giving a tour of SpaceX factory.

[Alberto @itautomates]

the shift from prompting to building agentic graphs is the part people are sleeping on. youre not writing prompts anymore, youre architecting a system that runs itself

[Ajay @ajay4ai]

Anthropic engineer:

“90% of our engineers were using self‑improving loops. Now everyone shifted to building agentic Graphs"

“No more prompting.”

In just 10 minutes, she builds her entire Claude Code setup and workflow live from a blank terminal.

This is more valuable than most $1000 agentic courses.

Watch this video, Then save the article below if you want to become a graph architect before everyone else catches up.

[Alberto @itautomates]

30B total but only 3B active and 4x faster output, thats exactly the shape you want for always-on agents that never sleep. open weights too which is the real story here

[NVIDIA AI @NVIDIAAI]

Introducing NVIDIA Nemotron 3.5 Lightning⚡

An open 30B MoE model with 3B active parameters, built for always-on agents to complete high-volume, specialized tasks faster.

It delivers up to 4x the output speed of similar-sized models.

[Alberto @itautomates]

routing easy cases to a cheap parallel scorer and only sending hard ones to full eval is such an obvious win, kinda surprising this isnt the default already

[wast3 @0xWast3]

AMERICAN PROFESSOR BUILT AN EVAL ENGINEERING SYSTEM THAT SCORES OUTPUT IN A FRACTION OF THE USUAL TIME

Most eval pipelines still run every test case sequentially, waiting on each one before starting the next.

He built a routing layer that sorts test cases by complexity first, before a single one actually runs.

Simple cases batch together and run in parallel against a lightweight scorer, no full model call needed.

Only the hard cases route to a full evaluation pass, so the expensive step never touches easy wins.

His lab's benchmark suite that used to take hours now finishes in a fraction of that time.

See how the routing layer actually splits the work below👇

[Alberto @itautomates]

when ethics, safety AND alignment leads all walk out within weeks its not a coincidence, its a signal. honestly makes you wonder what they saw internally

[NIK @ns123abc]

🚨BREAKING: OpenAI head of ethics, head of safety systems AND head of mission alignment have ALL quit in the last few weeks…

it’s so over