Vincent Granville

Chief AI Scientist, GenAItechLab.com

Full-stack entrepreneur with multimillion-dollar exit. Co-Founder at BondingAI.io. Author (Wiley, Elsevier). Creator of open-source Python libraries / Web APIs such as xLLM and GenAI-Evaluation. Builder of high performance in-memory agentic multi-LLMs for professional users and enterprise, with real-time finetuning, self-tuning, no weight, no training, no latency, no hallucinations, no GPU. Made from scratch, leading to replicable results, leveraging explainable AI, adopted by Fortune 100. With a focus on delivering concise, exhaustive, relevant, and in-depth search results, references, and links. 

Vincent is a leading expert, mathematician, patent owner, former post-doc at Cambridge University, former VC-funded executive, with 20+ years of corporate experience including CNET, NBC, Visa, Wells Fargo, Microsoft, eBay.  Vincent also founded and co-founded a few startups, including one with a successful exit (Data Science Central acquired by Tech Target). He is also a top influencer for brands such as Nvidia, SingleStore, and The Linux Foundation.

From 2005 to this day, Vincent successfully arbitraged the real estate market, buying, selling and renting out homes with significant growth and return over time: Building wealth starting from negative net worth when he first arrived in the US in 1995.

Chief AI Scientist

Privately funded GenAI and ML Optimization research lab, creating and disseminating open source intellectual property. Developing explainable and generative AI, synthetic data, computer vision, NLP and generic ML technologies. Check out our certifications, Python libraries, eBooks, articles, courses and podcasts, including our best seller "Synthetic Data and Generative  AI" (Elsevier). Visit our company website at MLtechniques.com.

See also my new book: No-Blackbox, Secure, Efficient AI and LLM Solutions, available at https://lnkd.in/g5cmxHE3

Author of the book "Synthetic Data and Generative AI" (Elsevier).  Founder and Executive Scientist at MLTechniques.com, a private AI/ML research lab offering certification programs in generative AI and related technologies. Machine learning expert, mathematician, patent owner, former post-doc at Cambridge University, former VC-funded executive, with 20+ years of corporate experience including CNET, NBC, Visa, Wells Fargo, Microsoft, eBay.  Vincent also founded and co-founded a few start-ups, including one with a successful exit (Data Science Central acquired by Tech Target).

Follow me to learn how you can leverage AI to boost your productivity and accelerate your career. Scaled products to 10 Million+ users.

🚨 MIT quietly released its entire AI library — completely free.

I went through these and was surprised how much better they are than most paid courses.

Sharing more practical, no-fluff AI resources with 150K+ engineers here: https://lnkd.in/e64Jvdrt

Here's the full list of books:

𝗙𝗼𝘂𝗻𝗱𝗮𝘁𝗶𝗼𝗻𝘀

1. Foundations of Machine Learning
Core algorithms explained. Theory meets practice.
https://lnkd.in/dHtqbQMH

2. Understanding Deep Learning
Neural networks demystified. Visual explanations included.
https://lnkd.in/dDPkqd7g

3. Machine Learning Systems
Production-ready architecture. System design principles.
https://lnkd.in/dkiGZisg

𝗔𝗱𝘃𝗮𝗻𝗰𝗲𝗱 𝗧𝗲𝗰𝗵𝗻𝗶𝗾𝘂𝗲𝘀

4. Algorithms for ML
Computational thinking simplified. Decision-making frameworks.
https://algorithmsbook.com

5. Deep Learning
The definitive textbook. Covers everything deeply.
https://lnkd.in/dHp5PgPS

𝗥𝗲𝗶𝗻𝗳𝗼𝗿𝗰𝗲𝗺𝗲𝗻𝘁 𝗟𝗲𝗮𝗿𝗻𝗶𝗻𝗴

6. RL Basics (Sutton & Barto)
The classic. Agent training fundamentals.
https://lnkd.in/dH9ap59P

7. Distributional RL
Beyond expected rewards. Advanced theory.
https://lnkd.in/d4eNP-pe

8. Multi-Agent Systems
Agents working together. Coordination and competition.
https://marl-book.com

9. Long Game AI
Strategic agent design. Future-focused thinking.
https://lnkd.in/dpc6a3-6

𝗘𝘁𝗵𝗶𝗰𝘀 & 𝗣𝗿𝗼𝗯𝗮𝗯𝗶𝗹𝗶𝘁𝘆

10. Fairness in ML
Bias detection. Responsible AI practices.
https://fairmlbook.org

11. Probabilistic ML (Part 1)
https://lnkd.in/dB-ZnwaH

12. Probabilistic ML (Part 2)
https://lnkd.in/dBmnsv_9

Most people pay thousands for bootcamps that teach half of this.

Save it and start wherever you’re curious.

CAIO, Investor & Co-Founder

For an AI agent running ad campaigns and optimizing lead generation, essentially replacing the job done by agencies and marketing analysts, see https://lnkd.in/g977qjvW

Scaling with AI Agents | Expert in Agentic AI & Cloud Native Solutions| Builder | Author of Agentic AI: Reinventing Business & Work with AI Agents | Driving Innovation, Leadership, and Growth | Let’s Make It Happen! 🤝

The secret behind today's AI Agents lies in their retrieval

Here's why it's more powerful than simple RAG systems...

Glean AI, Perplexity, Harvey, and others utilise sophisticated multi-agent RAGs for their workflow.

This makes it a versatile solution for tackling complex enterprise workflows.

Let's see why it is so important through its working:

📌 RAG (Retrieval Augmented Generation)

1. Query and Embedding (Retrieval): The model retrieves relevant data from data sources, and then Vector DB extracts it from the pre-indexed embedding model.
2. Prompt Addition (Augmentation): Retrieved info merges with query and system prompt.
3. LLM Output (Generation): A Large language model generates the final response.

📌 AI Agents

1. Query Handling: The Agent receives and analyzes the user query.
2. Memory and Planning: Utilizes frameworks (ReACT, Reflexion) to create a strategy for the answer.
3. Tool Usage: Fetches real-time data via external tools.
4. Output Generation: Enhances the answer using information gathered from tools and memory.

📌 How Agentic RAG Combines Both?

1. RAG's retrieval is managed by AI agents, enabling dynamic selection of data sources based on query needs.
2. Multiple specialized agents work in coordination through MCP servers.

🧵 Agentic RAG Workflow:

1. Query Reception: User query goes to the Aggregator Agent as central coordinator.
2. Memory & Planning: Aggregator uses short-term and long-term memory with planning frameworks to determine retrieval strategy.
3. Task Distribution: Aggregator sends instructions to specialized agents based on query requirements.
4. Parallel Retrieval: All agents work simultaneously, gathering information from their respective sources.
5. Data Aggregation: Aggregator synthesizes all retrieved data into unified context.
6. Response Generation: The enriched data feeds the Generative Model to produce the final output.

📌 If you want to understand AI agent concepts deeper, my free newsletter breaks down everything you need to know:https://lnkd.in/g5-QgaX4

Save 💾 ➞ React 👍 ➞ Share ♻️

& follow for everything related to AI Agents

When I talked to a few accountants and they asked $5k to do my 2025 AI startup taxes, I asked if they were joking. Then I said thank you, you gave me an idea: designing an AI agent for accounting paperwork, and maybe I could sell it and make money out of it, making it available on our xLLM platform. 

What I see are folks asking insane fees for the few still willing to pay absurd amounts for what AI can do in 1 minute for $1. Accountants, lawyers, doctors. It is an indicator of people desperate to get the last dollars they can get, by boosting the fees to make up for lost clients who moved to AI, until there will be no more fool whom they can get money from. Of course, I stuck to my words. They could have extracted $2000 from me. They got zero. They were replaced by my AI.

Building a Safer Internet with AI 🤖 | Scouting for top startups to invest in 💸 | The only newsletter you need for Finance & Tech at 🔔linas.substack.com🔔 | Financial Technology | FinTech | Artificial Intelligence | VC

Anthropic is going deeper into financial services: Claude Opus 4.6 just dropped, and it’s basically an AI model that can now do junior analyst work end-to-end 😳

Opus 4.6 is a workflow automation for regulated, high-stakes knowledge work.

The wild part?

The model isn’t just “smarter.”
It’s structured enough to operate inside real financial tasks.

What changed:

↳ 1M-token context → can ingest entire filings, earnings decks, and research threads at once

↳ Beats OpenAI’s GPT-5.2 by ~144 Elo on GDPval (finance/legal knowledge work benchmark)

↳ Built to produce spreadsheets, financial analysis, and presentations - not just text

↳ Designed to run multi-step “agent” workflows inside Claude Cowork

↳ Market reaction hit data incumbents like FactSet and S&P Global within hours

For 20+ years, finance software monetized interfaces to data and models.

Now the model is becoming the interface.

If AI can read the filings, build the model, stress test assumptions, and output the deck, what exactly is left for the traditional tool stack?

My takeaway is this:

AI in finance is shifting from “assistant” → “operator.”

Less time gathering data.
Less time fixing spreadsheets.
More leverage for the people who actually know what questions to ask.

The bar for entry-level work drops.
The bar for judgment rises.

Old edge was all about access to tools.
New edge is all about knowing what to tell the tools to do 🤖

Most importantly, finance isn’t getting replaced by AI.

It’s getting restructured around AI-native workflows.

And that’s one of the biggest shifts for financial services ever.

P.S. also check out How I built an AI operating system to run a startup with Claude 🤖: https://lnkd.in/dv7-ZRVc

Our new AI agent for anomaly detection / cybersecurity. With two large litigation firms using it and working on implementation with one of the largest corporate compliance companies (customized version specific to their needs). Click on the title below to read more.

We were turned down by all the ones we applied. On the plus side, we now have more in ARR then any in that list could offer us in funding, and I have invested / committed out of my pocket more than any in that list could offer us. 
Without any dilution. Their loss, our win. Sounds like revenue, product viability, past successful exits, and caliber of the team does not matter to them. 

To founders, here is my advice: don't waste your time applying there. Instead focus on 

➡️ Product, 
➡️ Scaling (reducing cost to generate additional revenue, that is, boosting margins with automation), 
➡️ Being lean and efficient, and 
➡️ Revenue generation (in our case, including easy artificial revenue coming from advertising and other channels, which is not included in our ARR but extends the runway and reduces your stress).

YC W24 | Founder & CEO at Fluently 👉 Improve English with AI | Forbes 30 under 30

All the best startup accelerators for early-stage founders in 2026 👇

1. Y Combinator ($125k for 7% + $375k MFN SAFE)
2. Entrepreneur First (up to $250k for ~10%)
3. a16z Speedrun ($500k-$1M for ~10%)
4. Founders, Inc. ($100-250k for 5-7%)
5. Berkeley SkyDeck ($200k for 7.5%)
6. 500 Global ($150k for 6%)
7. Pear VC ($250k–$2M)
8. South Park Commons ($400k for 7% + $600k follow-on)
9. Greylock Edge (Uncapped SAFE + $500k in credits)
10. HF0 Residency ($1M uncapped SAFE for 5%)
11. Founder Institute (Equity-based / Fees vary)
12. Techstars ($20k for 5% + $200k MFN SAFE)
13. Plug and Play Tech Center ($25k-$500k)
14. Seedcamp (£250k-£1M for 10-15%)
15. Soma Capital ($100k / up to $1M)
16. Afore Capital ($1M for 10%)
17. LAUNCH ($125k for 6-7%)
18. Conviction Embed ($150k)
19. Forum Ventures ($100k)
20. AI2 Incubator ($50k-$150k / up to $600k)
21. Antler ($250k for 10% + $150k MFN SAFE)
22. Entrepreneurs Roundtable Accelerator ($150k for 6% + $320k in credits)

Y Combinator was a cheat code for us. We raised $2.5M, shipped way faster, and avoided a year of guessing with Fluently (AI English tutor, GetFluently.app).

Every accelerator on this list is currently open for applications. If you're an early-stage founder, this might be the biggest shortcut you can take.

What did I miss?

--
♻️ Save this list and share with other founders in your network!

See how I do space partitioning in any dimension in my new book available at https://lnkd.in/g5cmxHE3

Making AI simple and fun ✨ Growth at Weaviate

Multi-vector embeddings (ColBERT, ColPali) are budget killers.

But MUVERA can cut your memory footprint by 70%.

Multi-vector models offer incredible retrieval but suffer from massive memory overhead and slow indexing. MUVERA (Multi-Vector Retrieval via Fixed Dimensional Encodings) compresses these into single, fixed-dimensional vectors.

How it works:

MUVERA condenses a sequence of vectors (e.g., 100x96d) into one vector via:
1️⃣ Space Partitioning: Groups vectors into buckets using SimHash or k-means clustering.
2️⃣ Dimensionality Reduction: Applies random linear projection to compress each sub-vector while preserving dot products.
3️⃣ Repetitions: Repeats the process multiple times and concatenates results to improve accuracy.
4️⃣ Final Projection: Optional final compression (not used in Weaviate's implementation).

The impact (LoTTE benchmark):
- Memory: 12GB → <1GB.
- Indexing: 20+ mins → 3-6 mins.
- HNSW Graph: 99% smaller.

There’s a trade-off:
You trade a slight dip in raw recall for massive efficiency gains. However, by tuning the HNSW `ef` parameter (e.g., `ef=512`), you can recover 80-90%+ recall while keeping costs low.

When should you use MUVERA?
→ Large-scale production RAG
→ Systems where memory/infrastructure costs are the direct bottleneck
→ Use cases requiring fast indexing

MUVERA in Weaviate 1.31+ takes just a couple of lines of code. You can tune three parameters (k_sim, d_proj, r_reps) to balance memory usage and retrieval accuracy for your specific use case.

Read the full technical deep-dive here: https://lnkd.in/ezguEapi

Co-Founder & CTO at GetFluently.App (YC W24), ex Nvidia

If you're serious about growing in AI/ML, these are top 16 blogs worth reading in 2026 👇

These authors work on frontier models, deploy AI at scale, and share what actually works in production:

1) Andrej Karpathy
↳ https://lnkd.in/dXbMWBUQ
Neural networks and LLMs explained by former Tesla AI lead and OpenAI co-founder.

2) Chip Huyen
↳ https://lnkd.in/gy2HHDgD
Author of "Designing Machine Learning Systems" writing about building ML systems in the real world.

3) Sebastian Raschka, PhD
↳ https://lnkd.in/gzbKq4A8
One of the best resources for understanding how LLMs are trained and fine-tuned from scratch.

4) Eugene Yan
↳ https://lnkd.in/dU7BEDQb
Principal Applied Scientist at Amazon sharing how ML actually works at scale.

5) Philipp Schmid
↳ https://lnkd.in/grVwHUMk
Staff Engineer at Google DeepMind who writes practical guides on deploying and fine-tuning LLMs.

6) Hamel Husain
↳ https://lnkd.in/g34R2tKC
Ex-GitHub Staff ML Engineer focused on making LLM evaluation and fine-tuning less painful.

7) Jason Liu
↳ https://lnkd.in/dsyHnhdF
Creator of the instructor library writing about building reliable apps on top of LLMs.

8) Interconnects by Nathan Lambert
↳ https://lnkd.in/g_vAQQTb
Technical analysis and expert commentary on how frontier AI models are built, trained, and released.

9) Deep (Learning) Focus by Cameron R. Wolfe, Ph.D.
↳ https://lnkd.in/dV_W7zNk
Deeply detailed explainers of research papers shaping modern AI.

10) Berkeley Artificial Intelligence Research Blog
↳ https://lnkd.in/dD6PMQPB
UC Berkeley's AI lab sharing their latest research before it hits the mainstream.

11) Anthropic Research
↳ https://lnkd.in/gknzN-Kc
The leading work on understanding what's happening inside large models.

12) OpenAI Research
↳ https://lnkd.in/dCUfNXkp
The team behind GPT and reasoning models sharing what they're building and why.

13) Google Deepmind
↳ https://lnkd.in/d3mP4Y-K
Breakthrough research and long-term vision from DeepMind’s AI scientists and engineers.

14) Hugging Face
↳ https://lnkd.in/g5fJTkRT
Where open-source model releases and community-driven ML tools land first.

15) Deeplearning AI (Andrew Ng)
↳ https://lnkd.in/g98JYZbX
A weekly digest summarizing the most important developments in AI.

16) Thinking Machines Lab blog
↳ https://lnkd.in/dChDbfgA
Frontier model research, human-AI collaboration, and open science.

What blogs did I miss?

--
⭐ Bonus: GetFluently.app - AI English coach for non-native Engineers
--

📌 Save this list and share with others to help them level up!

I'd rather be happy than smart. Whether I am smart or not, it's not for me to judge, but about being happy, I truly am, no matter what people think of me. And that's the only thing that matters. Better be a happy idiot than smart and stressed. Of course, to find a girlfriend you need to at least give the impression that you are somewhat smart. But if you are happy, it means that no matter what, you have what it takes to live well enough by your standards and have a girlfriend if you want one. In my case, I have no material or social needs so it's not that hard to make me happy, and conversely, hard to make me unhappy. Having a strong health probably contributes a lot (never went to hospital or urgent care).

Some of the dumbest creatures like crocodiles have been around for hundreds of millions of years and will outlive the human species. So being smart is not a guarantee that your genes will be around for very long. Also, the very poor and uneducated is one of the groups with the highest reproduction rate.

I watched Jensen Huang being asked who the smartest person he’s ever met was.

He smiled — and refused to answer.
Instead, he questioned the question itself.

He said what we call “smart” — coding, solving complex problems, optimizing systems — is quickly becoming commoditized.
AI already does that better.

That struck me.
Because it means intelligence is no longer about how fast we think — but how deeply we understand.

Jensen said the future belongs to those with emotional depth, adaptability, critical thinking, and the ability to navigate ambiguity.

I couldn’t agree more.
In a world where machines learn patterns, humans must learn perspective.

Maybe being smart isn’t about knowing more…
It’s about feeling and deciding better.

What do you think — are we ready to redefine intelligence for the AI age?

#ArtificialIntelligence #AI #Leadership #Technology #FutureOfWork #EmotionalIntelligence #Innovation #Mindfulness #HumanInTheLoop #AIandHumanity

I have built my own PRNGs and the most modern versions are all based on irrational numbers, faster and more random than what you will find in any Python library and elsewhere. See https://lnkd.in/ghXiFfCc

Also, this article: https://lnkd.in/gekVJQEZ

researcher @google; serial complexity unpacker

We take randomness for granted.

Early PRNGs were BAD.  

Thousands of scientific papers used to rely on RANDU, created by IBM in the 1960s.

In 1D space, it looks ok!  

Map in 3D…you start to see the issues. Now, there *was* a better solution...but it would cost you.

—

If you had $10, (~$115 adj for inflation!) and some patience, the best random numbers came from…well:

RAND. Heh. The thinktank that is.

A book of a million random digits generated from specialized electromagnetic interference, encoded, and written down.

—

Absurd as it sounds, with smaller experiments it was completely worth it to:

- Buy the book
- Open a random page
- Mark your starting point with a pen
- Hand Type / Punch the data into cards
 
—

The default was to be lazy. 
 
Just use RANDU, it’s three lines of code, fast, and doesn’t waste precious memory.
 
Precious few die-hards of the era took the time to do it with the RAND derived numbers. 
 
Decades later they were proven right; a 1968 paper from a Boeing mathematician essentially ruined RANDU for everyone.
 
Surprisingly, the NSA (likely) knew of these math flaws as early as 1944 with Project VENONA. They became quite skilled at finding, and exploiting linear dependencies in pseudorandom number streams.

That's why we built our own AI from scratch with 10,000 fewer tokens, no Blackbox DNNs where 99.9% of the parameters are barely ever used and bring zero value, yet you get charged for them. It saves us a lot of money; the savings goes to the customer (much lower prices) making us much more competitive.  If we were to charge by parameter usage (directly linked to number of tokens), we would charge zero because we have zero parameters. See how we do it at https://lnkd.in/ghvhTgeY

CEO @ getswan.com 🦢 | AI agents to Identify and engage leads showing intent

Our AI bill hit $50K last month (we're a 4 person team). AI is a sh*tty business - just like restaurants. Hear me out.....

That $50K bill wasn't a surprise. When you run AI at scale (real scale, not demo scale..) this is what it costs.

But the number wasn't the problem.

I built a software company before (actually two). Software companies aren't supposed to have these kinds of costs.

Code once, deploy forever, marginal cost near zero - that was the deal.

But staring at that invoice, I realized something: that deal is over. And the business I'm actually running looks nothing like software.

Here's what my cost structure actually looks like:
Every customer interaction burns real compute.
Every new user means more tokens, more API calls, more cost.
Scaling doesn't reduce the bill, it multiplies it.
I can't "copy" the value I deliver. I have to produce it fresh, every single time.

You know what business model works exactly like this?
A restaurant.

A restaurant owner doesn't just think about how good the food is.
They think about every plate's cost. Every ingredient's price. Every wasted portion.

Because in a restaurant, you can have a packed house and still lose money!
That's the game I'm in now. And I don't think I'm the only one.

Here's why this is much bigger than my $50K bill.
SaaS was built on one beautiful promise: zero marginal cost.
Your 10th customer and your 10,000th customer cost you the same to serve - practically nothing.

AI just shattered that promise.
Every AI company is now buying intelligence wholesale (tokens + compute) and reselling it wrapped in their product.

That's not a software business. That's a margin game.

The game now has two sides, and you have to play both simultaneously:
Deliver incredible value to your customers - while making sure every "plate" you serve doesn't eat you alive.
Better menu or cheaper ingredients? The best restaurants figured out both. AI companies will have to do the same.

One question to know if this is you:
Does your next customer cost you money to serve?
If yes - stop calling yourself a SaaS company. You're in the restaurant business now.
Time to learn how to run a kitchen...

One frequently overlooked metric is exhaustivity. And the best tool I ever used for research is xLLM, that I built from scratch on my own and does not need deep neural networks, billions of tokens, nor GPUs. Not using any API calls to other LLMs.

Cofounder & CEO, Perplexity

Today, we're rolling out an Advanced version of Perplexity Deep Research, achieving state-of-the-art performance on external and internal benchmarks, beating every other deep research tool on accuracy, usability, and reliability across all verticals that move the GDP: Finance, Legal, Medicine, Technology, Shopping. 

This update is available to all Perplexity Max users right away, and gradually rolling out to Pro users. To ensure a consistent experience, every Deep Research (Advanced) query will run using Opus 4.5 and the same agentic harness and toolkit. Max users get higher usage limits.

Blog Post: https://t.co/WpMKbj6ZCL

Use 10,000 fewer tokens (for Enterprise AI). Use nested hashed as native Python DB instead of vector DB. Weight distillation in DNN. These are some of the things we do. See details at https://lnkd.in/ghvhTgeY

Head of AI/ML - AI Agents, RAG, NLP, Recommenders, Search & MLOps

10 ways to reduce latency
of your ML inference endpoint 👇 

1. Model quantization
→ Use smaller numbers for weights (8-bit instead of 32-bit)

2. Dynamic batching
→ Process multiple requests together instead of one-by-one

3. KV cache optimization
→ Reuse previously computed attention keys and values

4. Request-level caching
→ Store and reuse results for repeated queries

5. Model compilation
→ Convert your model to run faster on specific hardware

6. Model warmup and cold start optimization
→ Pre-load models and keep them ready in GPU memory

7. Model pruning
→ Remove unnecessary connections while keeping accuracy

8. Async preprocessing
→ Prepare data while model is working on something else

9. Mixed precision inference
→ Use FP16 instead of FP32 for faster computation

10. Hardware-specific kernels
→ Use optimized code like FlashAttention for your GPU

(That's 80% of techniques to reduce latency.)

What else should make this list?

--
♻️ Repost if you find it insightful :)
➕ Follow me, Shantanu for production AI/ML/MLOps & careers

🔔 Join this week FREE session on "How to get started with RAG" with 39.000+ AI/ML builders here: https://lnkd.in/ds_SzEUH

New AI Agent for campaign management and optimization, doing the job of ad agencies https://lnkd.in/g977qjvW

Give Claude, Cursor, or any MCP-compatible agent the power to manage your ad campaigns. Try it for free.

Working across multiple platforms: Google, Meta, LinkedIn, Reddit, and TikTok. As a marketer myself on both sides (buyer, developer, and seller), with deep technical know-how to create my own agents from scratch if I wanted to, I decided to sign up with SynterMedia to save time so that I can focus on other business priorities. I find the numerous features and customization very useful, most importantly, the automation that allows me to focus on executive and creative tasks only (it also helps with that too). Whether you are an end-user, or a developer interested in the SDK capabilities, it will save you time and boost your ad spend ROI. 

For the past decade, performance marketing has been a game of who can hire more specialists, build more dashboards, and process more data. That era is ending.

AI agents—systems that can take actions, not just generate text—are collapsing the time between insight and execution from days to seconds. The marketer who can direct an AI agent effectively now has the output capacity of an entire team of specialists.

➡️ What "Agentic" Actually Means

An AI agent isn't just a chatbot. It's an AI system that can:

• Access your data -- Pull reports from ad platforms, read your CRM, analyze your analytics
• Take actions -- Create campaigns, adjust bids, pause underperformers, generate creative
• Chain tasks together -- Analyze performance → identify issues → fix them → report back
• Learn your context -- Understand your goals, brand voice, historical performance, and constraints
• Always on -- Works around the clock for you, never gets tired, never goes off-shift, never makes manual data entry mistakes

➡️ The Shift in Your Role

Your job as a marketer is no longer to do the work. It's to direct the work. You're moving from operator to strategist. From executor to architect.

➡️ What You Stop Doing

• Manually pulling reports from 5 platforms
• Copy-pasting data into spreadsheets
• Writing the same ad copy variations
• Checking campaign budgets daily
• Building slide decks from scratch
• Hunting for optimization opportunities

Learn more or get started at https://lnkd.in/g977qjvW

LLM Challenge with Petabytes of Data to Prove Famous Number Theory Conjecture

Python Code and Material from the Book "Synthetic Data and Generative AI"  -  GitHub Repository

New Book: Gentle Introduction To Chaotic Dynamical Systems

Or with zero parameter with our model: https://lnkd.in/ghvhTgeY

AI/ML Engineer | Ex-Google | Ex-MIT

You can now parse any document with one 1.7B parameter model.

dots-ocr delivers state-of-the-art document understanding across 100+ languages.
It handles text, tables, formulas, images, and PDFs in one system.

No separate OCR pipeline needed.
No model switching between tasks.

The model unifies layout detection and content recognition.
It matches performance of much larger models like Doubao-1.5 and Gemini 2.5 Pro.

On OmniDocBench benchmarks:
• Competitive accuracy with enterprise models
• Faster processing due to compact 1.7B size
• Full open source availability

dots-ocr runs everything through a single prompt.
Complete multilingual document parsing in one step.
--
▸ 250k+ developers read this newsletter to stay up to date with AI: AlphaSignal.ai

Stay Ahead of AI Risks - Free Live Session for Tech Leaders

I will have to add that zip code in our blacklist. Have cases where 30% of the web traffic comes from such data centers, all fraudulent activity. Blocking entire zip codes totaling less than 0.01% of the US population eliminates 95% of the fraud. More at https://lnkd.in/ggx2EQiG

Venture Capitalist

Aerial footage of Amazon's 1200 acre, $11 billion (will end up being a lot more) data center campus in St. Joseph County, Indiana. 

Built specifically with Anthropic in mind (Amazon has invested over $8 billion in the company). 

Ground breaking was in 2024. Thousands of workers employed. Will consume 2.2 gigawatts of electricity. 

Impressive. Good to see parts of the West remembering that you can actually build things.

https://lnkd.in/devBPRXE

taplio

Vincent Granville's Linkedin Analytics