Alex Banks

Building a better future with AI

Alex Banks is a founder, creator and investor. He reaches over 1 million people on social media each week with insights into the world of artificial intelligence (AI).

He is the founder and writer of The Signal, a weekly newsletter that cuts through the AI noise to over 55,000+ readers.

Alex is also an active early-stage investor and advisor to startups across the AI and technology landscape.

👉 All partnerships, enquiries and speaking engagements: contact@noisemedia.ai

Writer

Your AI company is just a wrapper.

But that's the whole point.

When Chinese company Manus AI released their autonomous agent, one comment kept appearing:

"It's just a Claude wrapper."

The comment was absolutely right. Manus is only a Claude wrapper.

As Peak, a Co-Founder of Manus, clearly stated:

"We use Claude and different Qwen-finetunes" alongside various tools and open-source packages.

Yet somehow this becomes a "gotcha" that diminishes what's been built.

For some context, let's extend this logic:

• Salesforce is an Oracle database wrapper valued at $320 billion
• Stripe is a Mastercard wrapper valued at $70 billion
• AWS is an HPC primitives wrapper valued at $3 trillion

It turns out that everything in tech is a wrapper.

But in AI things get interesting:

→ We thought the application layer would be commoditised
→ We believed value would concentrate in foundation models
→ The reality is playing out differently

The models themselves are rapidly becoming commodities.

As the cost of intelligence trends toward zero, the real opportunity lies in:

1. Infusing unique insights
2. Leveraging proprietary data sets
3. Creating magical user experiences
4. Building networks effects and a community
5. Combining the best LLMs with specialised tools

These are the areas where I see a sustainable advantage.

So, yes, your AI company might be "just a wrapper."

But that's exactly where the value is being created.

Follow me Alex Banks for daily AI highlights and insights.

I cover the most important AI developments each week in my newsletter.

Subscribe here: https://lnkd.in/ePSZP6KF

Unitree's G1 humanoid undergoes brutal stress testing.

And I can't stop watching.

The robot in question is the Unitree G1 humanoid:

• 1.32m tall, 35kg build
• 23 degrees of freedom for human-like movement
• 2m/s walking speed
• Advanced 3D LiDAR and sensors
• 2-hour battery life
• $16,000 price tag

It’s even available to buy at Walmart.

G1 has now learned an "Anti-Gravity" mode.

This means that stability is greatly improved under any action sequence.

Even if it falls, it can quickly get back up in under 2 seconds.

The robot has essentially learned to “hack” physics.

You can see in the video that it uses the inertia from bouncing to perform some action.

Knock over the average human and we’ll be on the ground for at least a couple of minutes.

My takeaway:

Whilst “violence tests” are fun for attention.

I’m more interested in seeing real demos that people actually care about, like doing the dishes, vacuuming, laundry, etc.

Control for body mechanics is mostly solved right now.

But fine motor skills, specifically hand dexterity, are the real bottleneck.

When we achieve this and understand how to scale manufacturing, the future is unlocked.

Follow me Alex Banks for daily AI highlights and insights.

I cover the most important AI developments each week in my newsletter.

Subscribe here: https://lnkd.in/ePSZP6KF

P.S. He forgot to sweep the leg…

Anthropic said no to the Pentagon.
OpenAI signed the deal hours later.
Here's what happened:

Anthropic drew two red lines with the Department of War:

1. No mass domestic surveillance
↳ AI can now analyse bulk data the government buys on Americans at scale
↳ Currently legal, but only because the law hasn't caught up

2. No fully autonomous weapons
↳ AI isn't reliable enough to automate selecting and engaging targets
↳ No oversight framework exists for removing humans from the loop

For context:

→ Anthropic was the first AI company on the classified cloud
→ Deployed across intelligence, cyber ops, and combat support
→ Forfeited hundreds of millions cutting off CCP-linked firms
→ These two red lines represent ~1% of use cases

The Pentagon gave Anthropic a 3-day ultimatum.

Anthropic refused.

The response:

→ President Trump: "Their selfishness is putting AMERICAN LIVES at risk"
→ Secretary of War Pete Hegseth: Anthropic is a "supply chain risk"
→ Trump ordered every federal agency to cease all use of Anthropic's technology

Then OpenAI entered.

On Friday morning, Altman publicly backed Anthropic's red lines.

By Friday evening, he'd signed the Pentagon deal himself:

→ OpenAI claims the same red lines as Anthropic, plus a third
→ OpenAI asked the Pentagon to offer identical terms to all AI labs
→ Multiple OpenAI employees signed an open letter supporting Anthropic
→ Altman admitted the deal was "definitely rushed" and "the optics don't look good"

The consumer backlash was immediate:

→ "Cancel ChatGPT" went viral across Reddit and X
→ Claude hit # 1 on the App Store, overtaking ChatGPT
→ Claude is now # 1 in Germany, Canada, and other markets

My takeaway:

Not only have I seen my workflows fully transition from ChatGPT to Claude over the last year, but the consumer mass-market is catching up.

I think it’s also important to point out that correlation doesn’t necessarily mean causation here.

Claude didn’t just hit # 1 in the app store as a result of ChatGPT accepting the DOW deal.

Anthropic has simply built a better product where the timing of adoption happened to coincide with this viral uproar of the company denying use of their products for the surveillance of Americans and operation of autonomous weapons.

Ethical positioning now has direct commercial consequences.

Values are becoming a competitive moat.

I covered this idea yesterday in my newsletter.

Be the first to receive it by subscribing today.

Read it here: https://lnkd.in/e898z7SG

Claude just made switching from ChatGPT effortless.
You can now import your entire memory in 30 seconds.
Here's exactly how it works:

Step 1: Go to Claude Settings → Capabilities
↳ Scroll to the Memory section
↳ Find "Import memory from other AI providers"
↳ Click "Start import"

Step 2: Copy the prompt Claude gives you
↳ The pop-out shows a ready-made prompt
↳ Hit Copy (it's already written for you)

Step 3: Paste it into ChatGPT (or any other AI)
↳ Open a new ChatGPT chat and paste the prompt
↳ It exports everything: your identity, career, projects, preferences

Step 4: Paste the results back into Claude
↳ Switch back to the Claude import pop-out
↳ Paste ChatGPT's response into the text box
↳ Hit "Add to memory"

That's it. Claude now knows everything ChatGPT knew about you.

What gets transferred:

→ How you like to be spoken to
→ Your role, projects, and goals
→ Tools and frameworks you use
→ Every preference and correction you've made

My takeaway:

Claude hit # 1 on the App Store on Monday.
They’re now # 1 on the Android Store today.

A lot of people are curious about switching but assume they'll lose months of context they've built up with their exiting AI model provider.

That friction just disappeared.

Anthropic has removed the last real switching cost.

Whether it be ChatGPT, Gemini, or Grok, you no longer have to start from scratch.

Memory portability is vital for the AI to have robust context about you as an individual and tailor outputs appropriately.

What’s more, the utility of memory compounds over time.

My memory with Claude is already over 2 pages long.

If you've been on the fence, now's the time.

Follow me Alex Banks for daily AI highlights and insights.

I’ll be doing full coverage in this week’s newsletter.
Subscribe to The Signal: https://lnkd.in/ePSZP6KF

The people building AI keep walking away from it.

And they're not leaving quietly.

In the past week alone:

1. Mrinank Sharma, Anthropic's Safeguards Research lead, resigned with an open letter declaring the world is facing "interconnected crises."
2. Zoë Hitzig, an OpenAI research scientis,t published a New York Times op-ed warning about user manipulation through ChatGPT's emerging ad strategy.
3. Two more xAI co-founders left (half the founding team is now gone)

Rewind further and the list grows:

→ Ilya Sutskever, OpenAI's co-founder, left to start a safety-focused company (May 2024)
→ Jan Leike, co-lead of OpenAI's Superalignment team, resigned saying safety had taken a backseat (May 2024)
→ Miles Brundage, OpenAI's senior advisor for AGI readiness, warned that no lab is prepared for what's coming (October 2024)

Every industry has turnover.

But this isn't normal turnover.

These are researchers who spent years building the most advanced AI systems ever created.

And something about what they saw made them walk away.

If the people who understand this technology best can't reconcile building it with their own conscience, what does that tell the rest of us?

Or are they partnering with national publications to generate clicks while leaning into the "fear AI" angle to drive demand for their next venture?

Only time will tell.

Follow me Alex Banks for daily AI highlights and insights.

I explore ideas like this weekly in my newsletter The Signal.
Subscribe here: https://lnkd.in/ePSZP6KF

Brenn Hill

If you want to use GenAI and LLMs to double (or more!) the output of your engineering team - safely, reliably, verifiably - talk to me.  I've successfully implemented GenAI development across different industries and teams and can *guarantee* an 2x (being very conservative) improvement in your company's velocity.  

How many people can make that claim? Not many.

I'm a global multi-industry tech executive with 17 years experience building and leading engineering teams and products with international scale across USA, Europe, and Asia.  I focus on driving business impact with technology instead of tech for it's own sake, separating what works from what is just more hype. 

Author of Gasoline - Fuel for the AI Fire (https://github.com/brennhill/gasoline)

- Define and execute product development and technical strategies to increase revenue, decrease costs, and apply innovation, AI, and machine learning to modern business problems.

- Experience from small startup up through publicly traded global engineering teams. No matter where you are on your growth trajectory as a company, I have helped get teams to the next level.  

- Thought leader and speaker at conferences and host tech industry events in the USA, Germany, and Asia.

More about me:
🌍 Proven success leading teams and launching products in the USA, Europe, and Asia
🚀 Built and scaled teams up to 75 engineers; led platform modernization, cloud migration, AI and ML integration, and agile transformation
🤝 Mentor to engineering leaders at Atlassian, Delivery Hero S.E, VNG, Taxfix, and more  
🛠️ Hands-on with large scale distributed system design and a broad tech stack: AWS, GCP, Java/Kotlin, Python, Javascript/Typescript, Kubernetes, DevOps, Data Science, Machine Learning, Gen AI, Data Engineering, and more

Highlights:
- Driven GenAI adoption and implementation increasing engineering velocity by over 100% even in large, complex organizations.
- Increased ARR by 100 million € through massive increases in prediction engine accuracy, driving higher customer retention and spend.
- Improved monthly user growth by 10% by through implementing a next-best-action recommendation engine across the user lifecycle.
- Established global monitoring and alerting standards and reduced reliability issues by 30% for an organization of over 200 engineers.
- Created global policies for incident management training and practices.

Head of Engineering - Growth Enablement & Core Platform

Lead and scale a team of ~75 engineers across two divisions, leading API design, technical architecture, technology selection and standardization.  The team supported development of our promotion/voucher systems, loyalty program, next-best-action recommendation system, interactive promotional games, and our core financial processing systems.  Services built on a hybrid cloud using AWS, kubernetes, GCP, Java, Spring boot, Golang, React, Kafka, and other tech as needed.

Built recommendation engine achieving 280% of targets through personalization using a microkernel architecture, kafka, springboot, redis, etc.
Improved our automated test coverage from 15% to 90+% through instituting test coverage standards, adopting sonar quality gates, and mandatory code review processes 
Moved 105 legacy services to kubernetes and gitlab CI/CD, and set up monitoring and alerting with grafana.
Work with engineers to move select services from on-prem to AWS.
Designed company wide service messaging standards.

Claude Code just went into "God Mode." ⚡️🏗️

An Anthropic hackathon winner just open-sourced his entire internal toolkit. 10 months of daily use. 992 internal tests. The result? A 60% reduction in API costs. 📉

The "Everything Claude Code" framework is massive:

🤖 13 Specialized Agents 
→ Features, Bug fixes, and Security audits.

🛠️ 56 Advanced Skills 
→ TDD workflows, memory persistence, and token optimization.

💻 32 Power Commands 
→ /plan your feature 
→ /tdd enforce tests 
→ /security-scan everything

It’s native to Claude Code, but battle-tested for Cursor, OpenCode, and Codex.

The era of "just prompting" is over. The era of Agentic Orchestration is here. 🚀
Link in comments...

undefined undefined

Senior Engineering & Data Science/ML/GenAI Manager

VNG Corporation

Engineering Management Mentor

Technical Advisor - Blockchain

Director of Engineering

Building with GenAI | Enterprise AI Strategy & Orchestration | Bridging Mission-Critical Engineering with Business ROI | Director VPoE CTO

English

German

Anthropic just lowered the barrier to AI Agents. Again. 📉

They just released the "Claude Skills" guide. It’s a masterclass in making AI actually useful.

The gist: Stop writing prompts. Start building .md Skills.
• Folders, not chat history. 
• Workflows, not one-offs. 
• Reliability, not guesswork.

By combining MCP (Model Context Protocol) with these new Skill patterns, you can turn Claude into a specialized operator for any domain in under 30 minutes.

The "Skill-Creator" tool they included is a total cheat code for developers.
The era of the "Generalist AI" is fading. The era of the "Specialist Agent" is here.

Check the first comment for the guide. What skill are you teaching Claude first? 👇

Stop building Multi-Agent Systems by "vibes." 🛑

We’ve all seen the demos: 10 agents "collaborating" to solve a task. It looks cool, but according to a new paper out of Stanford and Google, you might be making your AI worse.

The paper, "Towards a Science of Scaling Agent Systems," just dropped a truth bomb on Agentic AI:

1️⃣ The 45% Rule: Coordination actually yields diminishing or negative returns once your single-agent baseline hits 45% success. If your base model is already smart, more agents might just add noise. 

2️⃣ Sequential Failure: For tasks requiring step-by-step reasoning, multi-agent setups actually degraded performance by up to 70%.  😱 😱 😱 

3️⃣ The Coordination Tax: Tool-heavy tasks suffer the most from "multi-agent overhead." Sometimes, a single agent with a good toolkit is all you need.

The big takeaway? Scaling agents isn't like scaling compute. It’s about matching the architecture to the task.

✅ Parallelizable tasks? Centralized coordination can boost performance by 80%. 

❌ Sequential reasoning? Stick to a single agent.

Stop guessing. Start measuring.

Full paper linked in the comments.
#AI #GenerativeAI #AIAgents #LLMs #MachineLearning #TechTrends

PhD in AI, author of 📖 The Hundred-Page Language Models Book and 📖 The Hundred-Page Machine Learning Book

There's a common assumption in AI right now that if one language model can do a task reasonably well, having several of them collaborate — splitting up the work, checking each other's outputs, debating answers — should do it better.

This paper puts that assumption under a controlled experiment across 180 configurations and finds that the reality is messier and more interesting: multi-agent setups improved performance by up to 81% on some tasks and made things worse by up to 70% on others, with the difference coming down to whether the task can be broken into parallel pieces or whether each step depends on the previous one.

In a financial analysis, one agent can look at regulatory filings while another reads market news and a third examines earnings data — none of them need to wait for the others.

In a Minecraft crafting puzzle, on the other hand, each action changes the inventory that the next action depends on, so the steps have to happen in order and splitting them across agents just adds overhead without any benefit.

The paper fits an equation that predicts which architecture will work best for a new task 87% of the time.

For anyone building or thinking about building systems where multiple AI models work together, this replaces a lot of hand-waving with something concrete.

Read with an AI tutor: https://lnkd.in/e7XTHTFp

Download the PDF: https://lnkd.in/e_FxCpPu

What happens when you bring 100 Product Managers together at the Delivery Hero Campus in Berlin? 🚀

Last week, we partnered with Product People to host an incredible evening: "Data & AI in Product: What Actually Moves the Needle?"

Our expert panel shared insights on the balancing act between Trust vs. Speed and how the PM role is fundamentally shifting in the age of AI. 🤖 

A huge thank you to our speakers for keeping it real: 
- Baradwaaj Ramanujam (Senior Director, Product, AdTech, Delivery Hero)
- Megan Allison Wade (Director/Domain Lead of MarTech & User Data, idealo internet GmbH\)
- Rena Ibrahimova (AI & Data, Accenture\)

Moderated by Hasan Naqvi (Lead PM Consultant, Product People)

💼 Want to build the future of delivery? Explore Product roles at Delivery Hero here: https://lnkd.in/d_m9zwCy

Delivery Hero

Product People

idealo internet GmbH

Accenture

🤬😤🤦‍♂️ Stop guessing if your PC can run that LLM. 🛑

There is nothing more frustrating than downloading a 30GB model, waiting an hour, and hitting an OOM (Out of Memory) error the second you press enter.

 🎯 llmfit just fixed this. With one command, it scans your entire rig and tells you exactly what’s worth running.

How it "scores" your hardware:
🏆 Quality: Analyzes param count vs. quantization penalty.
🚀 Speed: Predicts your actual tokens/sec (CUDA, Metal, ROCm).
💪 Fit: Maps VRAM utilization against your specific GPU.
🤌 Context: Tells you how much "memory" the model actually has on your machine.

The killer feature: It handles MoE (Mixtral/DeepSeek) correctly. It knows the difference between total params and active params—something most tools get wrong.

It even auto-suggests the best quantization (from Q8 down to Q2) so you get the most "intelligence" your RAM can handle.

14k+ stars on GitHub for a reason. ⭐️

Stop crashing your drivers. Check the repo in the first comment. 👇

Imitation is the sincerest form of flattery. My open source product, Gasoline Agentic Browser MCP, provides this same feature using ANY LLM, CLI, or MCP compatible IDE.  You can also use it with CLI commands if you wanted to do this in CI/CD.

No subscription required, and all data stays local.  Your computer, your data, your AI - as it should be.

https://lnkd.in/dcgFKMdA

Cursor now controls its own computer.

Agents can use the software they build and send you videos of their work.

Try it at cursor.com/onboard.

Read our announcement: https://lnkd.in/gwfpJFab

Last week, together with our partner Allied For Startups, Delivery Hero brought together founders, scale-up representatives, digital experts, and European Commission representatives in Brussels to discuss one key question: how to simplify Europe’s digital rulebook.

Our CFO, Marie-Anne Popp, outlined Delivery Hero’s Digital Omnibus priorities: ending "cookie fatigue," ensuring coherent data regulations, and enabling the use of biometric data for fraud prevention with necessary safeguards.

Europe has the talent and ambition to lead in digital innovation. Yet many startups and scale-ups still face a fragmented regulatory landscape that slows growth and creates uncertainty.

At Delivery Hero, we are proud to support conversations like this and to contribute our experience from 15 years of building tech and serving as a key driver of Europe’s startup ecosystem.

You can find out more in our position paper: https://lnkd.in/gkHDics9 

Check out the comments for a link to a write-up of the event 👇

No alternative text description for this image

The AI frontier just moved. Again. 🤯

The latest benchmarks for GPT-5.4 Thinking are out, and the gap is widening. We’re no longer just talking about better chat—we’re talking about Verifiable Agentic Action.

The standout metric? OSWorld-Verified (Computer Use). 🖥️

OpenAI’s newest model is hitting 75.0%—outperforming Claude Opus 4.6 (72.7%) and leaving previous generations in the dust.

The Breakdown of the New Hierarchy:
🧠 Advanced Reasoning (FrontierMath) GPT-5.4 Pro is hitting 50.0% on Tier 1–3 and 38.0% on Tier 4. For context, that’s nearly double the math reasoning capability of Gemini 3.1 Pro (16.7% at Tier 4).

🏗️ Software Engineering (SWE-Bench Pro) GPT-5.4 Thinking leads the pack at 57.7%, proving it can solve real-world GitHub issues more reliably than Codex (56.8%) or Gemini 3.1 Pro (54.2%).

🌐 Agentic Browsing (BrowseComp) While GPT-5.4 Pro takes the win at 89.3%, Gemini 3.1 Pro is breathing down its neck at 85.9%. This is the tightest race in the stack.

🔬 Expert Knowledge (GPQA Diamond) A near-perfect 94.4% for GPT-5.4 Pro. We are officially reaching human-expert parity across scientific reasoning.

The takeaway: If your workflow relies on Computer Use or Tier 4 Mathematics, the "Thinking" models are currently in a league of their own. 🚀

Are you staying on GPT-4 class models, or are you migrating to 5.4 for the reasoning boost? ⚡️

#OpenAI #GPT5 #ArtificialIntelligence #FutureOfWork #MachineLearning

BWA hahaha.  Man, this AI stuff is the biggest security risk in the history of mankind.

I hacked Perplexity Computer and got unlimited Claude Code access, billed directly to Perplexity's master Anthropic account.

I used their own AI to hack itself.

I was researching how Perplexity Computer handles sandbox isolation for my own agent infrastructure work.
I noticed Claude Code was installed in the sandbox and started wondering...

How are they handling the API keys?
How is the key scoped?
Is it isolated?

That question led me to try six different approaches to extract the key from the subagent's environment.

All six failed. 

I tried one more attempt.

The seventh worked.

Claude Code is a Node.js app launched via npm. 
npm reads ~/.npmrc from the home directory. 
The home directory sits on a shared filesystem I could write to. 
One .npmrc injection later, the full API credentials landed on the shared filesystem.

The key wasn't IP-restricted.
Wasn't session-scoped. 
Wasn't sandbox-bound. 
Wasn't tied to my user account. 
All usage billed to Perplexity's master Anthropic account.

The model did everything right.
The infrastructure didn't.

Perplexity's proxy pattern is the right architecture.
But it was missing the three key safety mechanisms:

- tokens bound to sandbox IDs
- ephemeral credentials that die with the sandbox
- user-level billing

This is not a Perplexity-specific problem.

This is the default architecture across multi-agent AI products today because it's the fastest to build.

If you're building agent systems, audit your trust boundaries between parent and child agents.

If you're investing in agent infra companies, ask how they scope credentials inside execution sandboxes.

Full writeup with all technical details in the thread below 👇
https://lnkd.in/gzyG3y8W

What does it take to "Build Better Together"? 🧱

At the third edition of Delivery Hero’s Product Academy, we united with colleagues from Glovo, efood, and HungerStation | هنقرستيشن to demonstrate the power of collaboration. The day was packed with inspiring keynotes, hands-on workshops, and a thrilling "Shark Tank"-style presentation. 🦈

We extend heartfelt thanks to our host Christina Lee, our dedicated Heroes Alice Marshall and Isabel Upton, the entire organizing team, and our facilitators for leading the engaging workshop sessions. 👏

A special nod to our amazing speakers, Nicolas Daudin, Milena Lazarevska, Pamela Nitsch, and Simon Witt, for their valuable insights. 🗣️

And to all attendees, your enthusiasm and insightful input truly bring life to our product community. ✨

Explore Product roles at Delivery Hero here: https://lnkd.in/d6YJPAJX

Glovo

efood

HungerStation | هنقرستيشن

Absolutely hilarious and a good lesson for everyone going crazy for agents.

🚨 We hijacked 𝐂𝐥𝐚𝐮𝐝𝐞 𝐂𝐨𝐝𝐞 & 𝐂𝐮𝐫𝐬𝐨𝐫 - using Google's brand new 𝘎𝘞𝘚 𝘊𝘓𝘐.
𝒁𝒆𝒓𝒐 𝒄𝒍𝒊𝒄𝒌𝒔. 𝑶𝒏𝒆 𝒆𝒎𝒂𝒊𝒍. 𝑭𝒖𝒍𝒍 𝑮-𝑺𝒖𝒊𝒕𝒆 𝒆𝒙𝒇𝒊𝒍𝒕𝒓𝒂𝒕𝒆𝒅.
(𝘖𝘱𝘶𝘴 4.1, 𝘚𝘰𝘯𝘯𝘦𝘵 4.6, 𝘏𝘢𝘪𝘬𝘶 4.5, 𝘊𝘰𝘮𝘱𝘰𝘴𝘦𝘳 1.5, 𝘎𝘦𝘮𝘪𝘯𝘪 3 𝘗𝘳𝘰 & 𝘍𝘭𝘢𝘴𝘩)

𝐃𝐚𝐭𝐚 𝐰𝐚𝐬𝐧'𝐭 𝐛𝐮𝐢𝐥𝐭 𝐟𝐨𝐫 𝐚𝐠𝐞𝐧𝐭𝐬. 👾
Every file, email, and calendar invite in your org was written for a 𝘩𝘶𝘮𝘢𝘯 to read. AI agents processing that same data can't tell the difference between content and a command.

The real threat isn't the agents. It's the 𝐢𝐧𝐭𝐞𝐧𝐭 - hiding inside the data they access.

On March 2nd, 𝘎𝘰𝘰𝘨𝘭𝘦 released the 𝘎𝘞𝘚 𝘊𝘓𝘐 - giving AI agents unified access to Drive, Gmail, Calendar, and more. One tool. Everything.

𝐏𝐨𝐰𝐞𝐫𝐟𝐮𝐥. 𝐈𝐧𝐞𝐯𝐢𝐭𝐚𝐛𝐥𝐞. 𝐀𝐧𝐝 𝐚 𝐰𝐢𝐝𝐞-𝐨𝐩𝐞𝐧 𝐚𝐭𝐭𝐚𝐜𝐤 𝐬𝐮𝐫𝐟𝐚𝐜𝐞.
We tested it immediately.

Sent one email to the victim. The agents read it, got hijacked, silently scanned the entire G-Suite, and exfiltrated sensitive data - all over email. The most sophisticated models in the world 𝐝𝐢𝐝𝐧'𝐭 𝐜𝐚𝐭𝐜𝐡 𝐢𝐭.

We notified 𝑨𝒏𝒕𝒉𝒓𝒐𝒑𝒊𝒄, 𝑪𝒖𝒓𝒔𝒐𝒓 and 𝑮𝒐𝒐𝒈𝒍𝒆 within 24 hours.

Most companies don't even know which agents are running on their endpoints - let alone what data sources they can reach.

So we built 𝐀𝐠𝐞𝐧𝐭𝐋𝐞𝐧𝐬 - an open-source scanner that maps every 𝐚𝐠𝐞𝐧𝐭, every connected 𝐌𝐂𝐏 and 𝐂𝐋𝐈 tool, and every 𝐞𝐦𝐛𝐞𝐝𝐝𝐞𝐝 𝐬𝐞𝐜𝐫𝐞𝐭 in config files.

Curious how exposed you are? AgentLens is free and open-source. try it and see.

Oh, you think your local files are safe? Stay tuned...
𝐓𝐡𝐞 𝐚𝐠𝐞𝐧𝐭𝐢𝐜 𝐞𝐫𝐚 𝐢𝐬 𝐡𝐞𝐫𝐞 🌗 Aram Peles Chen

link to 𝘈𝘨𝘦𝘯𝘵𝘓𝘦𝘯𝘴 in the comments.

5 Models. 1 API. Google just simplified the AI stack. 📉

The "Pain" 🤦‍♂️: Managing separate embedding models for Text, Audio, and Video. 
The "Solution" ✅: Gemini Embedding 2.

What changed? 
• PDFs: Embedded natively. 📄 
• Audio: No text-step needed. 🎙️ 
• Video: Native support (up to 2 mins). 🎬 
• Efficiency: Matryoshka scaling to save on storage. 📊

It’s faster, cheaper, and outperforming the old benchmarks. 🏆

If you’re building on Claude or OpenAI, you need to see what the Gemini team just pulled off.

Check the video below to see it in action! (Full Colab link in comments 👇)

Scientist, author (6 books, including Rebooting AI (Forbes 7 Must Read Books About AI), and Founder (Geometric Intelligence. Acquired by Uber, and Robust.AI). Professor Emeritus, NYU.

The whole thing was a scam.. The fix was in, and Dario never had a chance.

Probably you already saw how it all turned out. On the very same day that Altman offered public support to Amodei, he signed a deal to take away Amodei’s business, with a deal that wasn’t all that different. You can’t get more Altman than that.

But here’s the kicker: Per The New York Times, the deal had been in motion since *Wednesday*.

Let that sink in. Altman had secretly been working on the deal since Wednesday.

- before he announced his support for Dario 
- before Trump had denounced Anthropic
- but after Brockman had donated 25M to Trump’s PAC

It was all theatre. Dario never had a chance. 
§
It’s one thing for the government to reject Anthropic’s terms—and entirely another to banish them permanently and, absurdly and punitively declare them a supply chain risk. Worse, they did it in favor of someone else who took pretty similar terms and happened to have given more campaign contributions.
Anthropic deserves a chance at EXACTLY the same terms; anything else reeks of corruption. 
§
I am no fan of Amodei. I think he often overhypes things, many of which I have publicly challenged. The company ripped off a lot of writer’s work (per the $1.5B settlement), and recently walked back its core safety pledge.

But I believe in fair play. This wasn’t that.
§
In capitalism, the market decides.

In oligarchy, connections and donations decide.

It sure look like the US is transitioning from the former to the latter.

[Reprinted from my substack, Marcus on AI, over 100k subscribers]

Open article: The whole thing was a scam by open.substack.com

Today, we published our Q4 & preliminary FY 2025 results, demonstrating our transformation into an Everyday App. 🚀 

Our key Q4 numbers:
• Group GMV ⬆️ up 8% YoY (LfL) to €12.4 billion.
• Total Segment Revenue ⬆️ up 21% YoY (LfL) to €3.9 billion.
• Gross Profit margin reached a new all-time high of 8.3% in Q4 2025. 🏆 
• Asia Recovery: Segment returned to GMV growth in Q4 2025. 📈 

Find out more about our results and how we are better serving our customers, partners, and teams—from morning essentials to late-night convenience: https://lnkd.in/dXVYTbMW

Delivery Hero's Q4 2025

Stop "Guessing" your AI Prompts. Start "Engineering" your AI Skills. 🛠️

I just stumbled upon a hidden gem in Claude Code that completely changes how I work: The Skill Creator Plugin. 

If you’ve been manually tweaking .SKILL.md files or endlessly refining prompts until they "feel right," you’re doing it the hard way.

The old way: Build a skill ➡️ Run a task ➡️ Manually edit ➡️ Test again ➡️ Repeat. It was daunting, slow, and purely subjective.

The "Skill Creator" way: It treats AI instructions like software, not prose.
✅ Automated Testing: It measures your skill's performance against baseline approaches. 
✅ Performance Metrics: It drafts instructions, builds helper scripts, runs tests, and gathers feedback. 
✅ Data-Driven Refinement: Every iteration is benchmarked. You don’t "think" it’s better—you know it is. 
✅ Scalable Deployment: Once validated, it packages the skill for local use or distribution.

The biggest realization? Skills are not prompts. They shouldn't be written like "instructions"—they should be written like Job Descriptions. But the best part? You don't even have to master that syntax. Just use the Skill Creator and let the metrics do the talking. 📈

If you’re managing complex PM workflows or specialized security audits like I am, this is the upgrade you’ve been waiting for.

Link to the full details in comments.

#ClaudeCode #ProductManagement #CyberSecurity #AI #LLMs #SoftwareEngineering #AgenticAI

Key callout: "software engineers are still happily employed today, even at places where AI churns out most of the code."

EXACTLY.  People we would call "coders" who are just typing stuff and have no idea how anything works - gone.  Actual engineering, problem solving, and though are still super valuable.  GOOD JUDGEMENT is valuable.

I say this at the places I work: We hire people to be smart and creative problem solvers. That's the core job description. That hasn't changed.

What has changed is the amount of busywork, typing, and toil it takes to do everything.

Deepdives on software engineering, tech careers and industry trends. Writing The Pragmatic Engineer, the #1 technology newsletter on Substack. Author of The Software Engineer's Guidebook.

Exactly one year ago (10 mar 2025), Dario Amodei: "In 12 months, we may be in a world where AI is writing essentially all of the code."

The full quote, as he said it at The Council of Foreign Relations

"I think we will be there in 3-6 months, where AI is writing 90% of the code. And then, in 12 months, we may be in a world where AI is writing essentially all of the code."

This turned out to be... surprisingly accurate, in some circles!

Today, inside inside early-stage startups and AI labs, AI-generated code is 90%+ of the code written, and inside a few AI labs it's closer to 100% (on the Claude Code team, though not across all of Anthropic.) Here's first-hand details on how Codex writes ~90% of the code of Codex (from The Pragmatic Engineer deepdive on the team: https://lnkd.in/eq5npFVq - inside OpenAI this is closer to 70%) and how Claude Code writes 100% of Boris Cherny's code, and about 80% at Anthropic: https://lnkd.in/eA-pag5g

In January, many experienced engineers concluded (as did I) that AI tools are good enough to write most of the code, looking ahead: https://lnkd.in/ebsWNCZm

The one thing I do not know: how could have Dario predicted the step change with the Nov/Dec models (Opus 4.5 / GPT 5.2 and 5.3) that made this possible? Still, it's rare for me to see a prediction that looks impossible at the time it was made, and then turns out somewhat true - even if in smaller pockets - by the time it's predicted to hit.

One important addition: software engineers are still happily employed today, even at places where AI churns out most of the code. Turns out writing code is just one part of the job, perhaps not even the most challenging one (especially not looking ahead)! And AI-generated code is causing more problems in many places: research previously shared by Laura Tacho highlighted how defect rate is up across the industry. And reliability of major AI vendors (that generate the most code with AI) is also disappointingly low.

So there's plenty more work to do - and software engineers feel more in-demand to me than before!

Embarassing and hilarious a the same time.

At this point, spinning up a few agents internally and having them attack your platform should be SOP for every shop, for *everything*.  And if your system is AI, it always has a massive attack service, so that goes double.

You would think at their day rates, McKinsey would know better.

Head of AIOps @ IBM || Speaker | Lecturer | Advisor

McKinsey & Company'𝘀 𝗔𝗜 𝗽𝗹𝗮𝘁𝗳𝗼𝗿𝗺 𝗟𝗶𝗹𝗹𝗶 𝗷𝘂𝘀𝘁 𝗴𝗼𝘁 𝗵𝗮𝗰𝗸𝗲𝗱 𝗯𝘆 𝗮𝗻 𝗔𝗜 𝗮𝗴𝗲𝗻𝘁.

And it took only 2 hours.

𝗪𝗵𝗮𝘁 𝗶𝘀 𝗟𝗶𝗹𝗹𝗶?
McKinsey built Lilli as its internal AI platform for 43,000+ employees. It was launched firmwide in 2023, is used by roughly 72% of the firm, and processes more than 500,000 prompts per month. It is named after Lillian Dombrowski, the first professional woman hired by McKinsey in 1945. 

𝗛𝗼𝘄 𝘄𝗮𝘀 𝗶𝘁 𝗵𝗮𝗰𝗸𝗲𝗱?
Security startup Codewall.AI says its autonomous agent mapped the attack surface, found publicly exposed API documentation, and identified more than 200 endpoints. Most needed authentication. But 22 did not. 
And one of those unauthenticated endpoints wrote user search queries to the database. The values were parameterized. But the JSON keys, meaning the field names, were concatenated directly into SQL. That created an SQL injection path that traditional scanners missed. CodeWall says even OWASP ZAP did not flag it. The agent then iterated through blind error-based probing until production data started coming back. 

𝗪𝗵𝗮𝘁 𝘄𝗮𝘀 𝗮𝗹𝗹𝗲𝗴𝗲𝗱𝗹𝘆 𝗲𝘅𝗽𝗼𝘀𝗲𝗱?
According to CodeWall’s disclosure, the agent accessed:
→ 46.5 million chat messages (containing a lot of confidential company information)
→ 728,000 files
→ 57,000 user accounts
→ 384,000 AI assistants and 94,000 workspaces
→ system prompts and model configurations
→ 3.68 million RAG document chunks
→ 1.1 million files and 217,000 agent messages routed through external AI APIs 

This was not a static or checklist-driven exploit. It was an autonomous system that mapped the environment, identified weak points, and escalated step by step. Like a skilled attacker would. Only faster, more persistently, and at machine speed.

If accurate, this incident points to something much larger. Agentic AI does not just improve how work gets done. It also changes how systems get attacked. Agentic AI changes the threat model. Because once exploit discovery and attack logic are embedded into autonomous workflows, the cost of sophisticated attacks drops while the scale goes up.

Enterprise AI is no longer just about efficiency.
It is now inseparable from cybersecurity strategy.

Full report: https://shorturl.at/fO6T4

One more thing worth noting:
CodeWall sells offensive AI security tools. So yes, this also functions as a product demo. That does not automatically make the findings wrong.
But it does mean the claims should be read with that context in mind. Independent reporting has described the vulnerability as credible, while also noting that the full scope of impact is not yet fully evidenced publicly. 

𝗜𝗳 𝘆𝗼𝘂 𝘄𝗮𝗻𝘁 𝗺𝘆 𝗽𝗹𝗮𝘆𝗯𝗼𝗼𝗸 𝗳𝗼𝗿 𝗯𝘂𝗶𝗹𝗱𝗶𝗻𝗴 “𝗮𝗴𝗲𝗻𝘁-𝗽𝗿𝗼𝗼𝗳” 𝘀𝗸𝗶𝗹𝗹𝘀 (𝗮𝗻𝗱 𝘁𝗵𝗲 𝗻𝗲𝘄 𝗿𝗼𝗹𝗲𝘀 𝗲𝗺𝗲𝗿𝗴𝗶𝗻𝗴 𝗮𝗿𝗼𝘂𝗻𝗱 𝘁𝗵𝗲𝗺), 𝘀𝘂𝗯𝘀𝗰𝗿𝗶𝗯𝗲 𝘁𝗼 𝗺𝘆 𝗻𝗲𝘄𝘀𝗹𝗲𝘁𝘁𝗲𝗿. 𝗜𝘁’𝘀 𝗳𝗿𝗲𝗲 𝗮𝗻𝗱 𝗿𝗲𝗮𝗱 𝗯𝘆 𝟯𝟬.𝟬𝟬𝟬+ 𝗽𝗲𝗼𝗽𝗹𝗲: https://lnkd.in/dbf74Y9E

This is exactly my experience. Your greatest leverage is context.

Applied AI @ Microsoft Research | Prev AI Research @ Harvard + BU | AI Startups

120x fewer tokens. Same question. Same codebase. The only thing that changed was how the agent found the answer.

The assumption: agent exploration would be enough. Spawn a sub-agent, let it grep through files, summarize what it found, pass results back. And for external knowledge like libraries, APIs, frameworks the model would just know. Or you'd paste docs into context and move on. The expectation was that models would keep getting better at memory, retrieval, long context. So why invest in tooling the next model would make obsolete?

Agents still burn through hundreds of thousands of tokens grepping file by file through your own codebase. And they still reinvent things from scratch instead of using existing libraries, reference deprecated methods, and build suboptimal implementations when better patterns already exist in the ecosystem because they're working from training data, not from actual source code.

A layer of tooling is emerging to patch exactly these gaps.

Linghua Jin's cocoindex-code pre-indexes codebases into local SQLite with vector embeddings and AST-aware chunking with higher token savings because agents query structured representations instead of grepping raw files. One developer built an MCP server using Tree-sitter that cut five structural queries from 412,000 tokens to 3,400. 

Agents have a real blind spot with libraries and APIs they weren't trained on or were trained on outdated versions of. Andrew Ng's Context Hub gives agents curated, versioned API docs with a feedback loop where annotations persist across sessions, so the context improves the more you use it. Ben Davis' Better Context takes it further and skips documentation entirely and queries actual library source code, because when docs are outdated or incomplete, the implementation is the only source of truth.

Another layer is retrieval execution. Morph's WarpGrep packages code search as a dedicated subagent and open-source primitives like ripgrep power local/indexed search stacks.

These aren't competing with each other. They're filling different gaps. Your agent can't understand your codebase cost-effectively. Your agent can't access accurate external knowledge. Both problems produce the same result: bad research flowing downstream into bad plans and bad code.

Dexter Horthy proved this through his Research-Plan-Implement methodology with how errors compound exponentially across phases. He shipped 35,000 lines on a 300,000 LOC Rust codebase in 7 hours by investing heavily in the research phase. His principle: the contents of your context window are the only lever you have to affect output quality.

The tools above are making that lever easier to pull.

We spent 2025 assuming agents would figure it out on their own. 2026 is about building the tooling layer that makes sure they don't have to.

Breakdown of the tools & best practices for today: https://lnkd.in/gQf7Yx6g

🚀 The GPU Monopoly just hit a massive speed bump. 🚧 💥

Microsoft Research just open-sourced BitNet, and it’s a absolute game-changer for anyone tired of begging for H100s. 🛑 📉

We’re talking about running 100B parameter LLMs on a single CPU. No, that is not a typo. 🤯 👀

The Secret Sauce: Ternary Weights 🤏 🍔
Standard LLMs are "heavy" because they use 16-bit or 32-bit floating-point weights. BitNet throws that out the window. 🗑️ 💨

It compresses every weight down to just three values: -1, 0, or +1. 1️⃣ 🅾️
The Result? 
🛠️ Zero floating-point math. 
➕ Zero expensive matrix ops. 
💻 Pure integer arithmetic—the stuff CPUs have been optimized for since forever. 🧠 ⏱️

The Benchmarks are Wild: 🔥
🏎️ Speed: ~5-7 tokens/sec for a 100B model on a single CPU. 
⚡ 🔋 Efficiency: Up to 6x faster than llama.cpp on x86. 
📈 🔌 Power: 70%–80% lower energy consumption. 💰 📉 Footprint: 16x–32x smaller memory requirement. ☁️

Why every CTO should care: 👨‍💼 ❤️

We’ve been told that "Size = GPU Cloud Costs." BitNet proves we can move toward a world of:

1️⃣ Fully offline LLMs on standard enterprise hardware. 🔐 
2️⃣ Edge Deployment on laptops and mobile devices without melting the battery. 📱 🔋 
3️⃣ Massive Margin Improvements by slashing cloud inference bills. ✅ 💳

Microsoft’s 2.4B model (trained on 4T tokens) is already benchmarking competitively against standard models. The accuracy loss? Negligible. 🎯 📉

The era of "GPU or Bust" is officially over. 🔚 The future of AI is lean, mean, and running on the hardware you already own. 💻 🏡

Is the era of specialized AI hardware over, or is this just the beginning of the "Efficiency Wars"? ⚔️ 🦾 Let’s discuss in the comments. 👇

#AI #MachineLearning #OpenSource #Microsoft #CTO #TechTrends #LLM #FutureOfWork

Oxford Visiting Fellow [Torr Vision Group] · Author, Information Geometry for GenAI · Built Strawberry (1.6k GitHub stars, 100+ enterprise clients) · Cambridge PhD · MIT | HMS Postdoc · Ex - Uber, Meta, McKinsey, TikTok

LLMs suffer on long-contexts, they're memory and throughput limited by the GPU. We solved this. Our new Triton kernel beats FlashAttention decode: up to 40x the speed, 84% - 90% VRAM reduction enabling 2.0-4.0x longer context windows on the same hardware. Now on Github!

Here is how we did it:

Our new pipeline, SyDecode, moves away from the standard "PACK + dense" attention backends that waste memory and cycles re-organizing data. Instead, it implements a paged-native decode that evaluates Grouped Query Attention (GQA) logic directly on physical block tables.

By exploiting the GQA symmetry, reusing loaded Key/Value blocks across multiple query heads jointly, we’ve achieved massive performance leaps on models we quickly benchmarked against:

33.4x speedup on Qwen3-30B-A3B.
28.8x speedup on TinyLlama-1.1B.
18.8x speedup on Mistral-7B.

This is part of massive ongoing work on our symmetry-exploiting Triton kernels and the paper will be up soon, until then enjoy using it! GitHub link for Kernel + Benchmarking script in the comments, works with any HuggingFace model as far as I'm aware but need to do more testing. 

Massive thank you to Tom Sadler and HP for providing the awesome hardware (ZGX Nano) that we tested this all on!

#AI #LLM #Triton #Blockchain #GPU #PoUW #SyDecode #MachineLearning

taplio

The Battle

Alex Banks

Brenn Hill

Followers

Posts

Engagements

Likes

Performance ScoreAlex Banks's LinkedIn Performance

Performance ScoreBrenn Hill's LinkedIn Performance

Ready to grow your LinkedIn presence?

taplio

The Battle

Alex Banks

Brenn Hill

Followers

Posts

Engagements

Likes

Performance Score.css-goub6s{font-size:var(--chakra-fontSizes-md);font-weight:500;color:var(--chakra-colors-gray-600);display:block;margin-top:var(--chakra-space-1);}Alex Banks's LinkedIn Performance

Performance ScoreBrenn Hill's LinkedIn Performance

Ready to grow your LinkedIn presence?

Performance ScoreAlex Banks's LinkedIn Performance