1082
Building https://alexvarga.xyz
18 months ago before Anthropic and ChatGPT we shipped agents with embedded UI components and 2-click MCP connectors
Today I decided to open source VeyraX - OpenRouter for tools
- 80+ integrations and thousands of MCP servers
- Embedded UI components for agents
- Flows - agent-powered multi-step workflow
MIT, self-hosted
GitHub
Why fundraising in the Bay Area is crazy.
To pre-train a good LLM, you need around 10,000 GPUs at $2/hr.
It’s around $175M/year.
And a research team has around 2 attempts for the year.
So you need to raise at least $200M.
If you want to train a smaller LLM, at least $10-20M
Jumping back to success and failures, I see a pattern that great companies require:
(a) Hardcore distribution
(b) Exceptional R&D (usually 3+ years)
And everytime #1 important factor is a team
…and luck
Interesting that in America there is a trend of Russian math
https://www.youtube.com/watch?v=5zCJZqeW4tY
Unethical startup idea:
You send your app, and AI decides what’s economically better: hack you or ask you to pay for bug fixes
Watched 2001: A Space Odyssey.
This is one of the highest-rated films, and according to most film directors, it may even be the most important film in the entire cinema industry.
More remarkably, it was scripted based on Arthur C. Clarke’s book. He was a person who predicted over 90% of the technology trends we have today, including longevity, computers, the internet, digital nomadism, AI, space travel, and other ideas. To understand how powerful his predictions were, I highly recommend watching his 1964 predictions about the future, which we would call the present today.
Stanley Kubrick made a genius film using Arthur’s 1948 book The Sentinel, and I highly recommend watching the film, which showcases two geniuses collaborating.
Another great sci-fi writer, Isaac Asimov, made equally impressive predictions in his books, especially regarding robotics, AI, and space travel.
I am impressed by how sci-fi books inspire the greatest entrepreneurs to build the future, and this is how science fiction eventually predicts the world we live in today.
I was wondering: Asimov, Clarke, and others lived in the previous century. It is not cool to hear about the future from information more than 50 years old.
So, first, I read Jacques Fresco — an engineer and futurist who researched for most of his life what it would be like to live in a world with no money. I like some of his ideas, but he passed away in 2017 and didn’t see much of the AI boom, despite predicting it already in the 1990s.
So, the easiest way to find really great futurists who are still alive and study the future is simply by asking ChatGPT.
One recommendation is Kevin Kelly, founder of Wired — one of the well-known tech media outlets.
Kevin predicted bitcoins in the 90s, distributed systems, and commoditized AI decades before they actually happened.
His latest predictions are:
- Robots will lag behind AI. This is mostly due to hardware limitations, especially because they require a lot of energy, and having a lot of energy at home is like having a mini bomb; also, there are many requirements for power systems, and robots need better world models, manipulation skills, and continuous learning.
- Emotional AI, where humans will spend most of their day with digital companions that can show empathy, love, anger, hate, and other feelings.
- No AGI. Instead, a swarm of AIs and robots, each of which can solve a small task well. Think of it as specialized aliens with godlike intelligence that can help humans do certain things.
- Bots as economic units. As the human population declines in some countries, bots can help keep production lines running.
- The Human-AI relationship will reach a new level. We will have more companions, and we will also pay more for real people serving us. Think of an AI therapist versus a real therapist. As AI gets better at showing emotions, it will no longer be weird to have AI friends and spend time in such a mixed company of real people and AI aliens.
One thing I was surprised by in his vision: the AI revolution is not at the beginning; in 2026, it has not even started. This is how he explains what we need to expect for the next 20-30 years.
I watched his prediction video but also inspired to read some of his works
I love how they force memes with Perelman!
Читать полностью…
Updated website a little bit, now you can steal our hero video because it is made with varg.ai
I also discovered too little use cases covered and added a bunch right on the frontpage.
With varg you can create videos from 5 seconds to 17 minutes (and I guess longer, just I never tried make it longer)
The best result is you want to use MCP server with Fable 5.1 that is the only good way how I suggest to use AI video models today
Played a little bit with H3 Max and made a 17 minute podcast video about Omega 3
It took me <30 minutes to make, and $40 total
https://x.com/vargastartup/status/2094901550867181837
A lot of excitement about AI in Math
Yesterday, someone found a small elliptic curve with rank >=30
y² + xy = x³ − 201769035260418549083594900060734240952308696994802735114305555 x + 1151107939141058565733479426024323225135665982951300586808823640527729578307228357301072889377
The idea:
- There is a nerd competition to find the smallest elliptic curve (y^2+axy = x^3 + bx) with the highest rank
- Elliptic curve rank = how many independent points you need to generate all its rational points
- Generating a rational point means there is a “sum” function on the elliptic curve that generates new points using existing ones
So basically, it’s like a competition for “how to write a program using the least amount of code,” but for mathematicians.
The result itself is completely pointless, because no one cares about high ranks, and it has no practical applications.
But it feels like a great example of the kind of research AI should eat alive.
Our brain is great at summing numbers (2+2=4), but bad at summing numbers on elliptic curves.
However, the problem is extremely formalizable
We can brute force and verify using AI because it’s all verifiable objects: the sum on curves is well defined
Rank 30 itself probably doesn’t matter
Every cinema is identical up to automorphism
Читать полностью…
I see no reason to use real people for podcasts 🤪️️️️
Seedance is just so good
I highly recommend to watch Dmitri Dolgov (Waymo Co-CEO) talk at Y-Combinator. He works at Google for over 20 years, and all of the time he developed Waymo, grew to Co-CEO role.
• he explains how Waymo - AI self driving, the only real mass adopted consumer artificial intelligence working
• He explains why software allows us to ship 99.9% of things, but each next 9 is ten times harder to achieve.
For example, for Waymo they need to be safety at least of 99.9999%; and it is ~1000x harder than 99.9%
That’s why it took them 18 months to ship a demo, and 20 years (!) to be live on the market (400 more months!)
https://www.youtube.com/watch?v=Gp4zrV3-6N8
Gym in the building is a luxury you get used to in San Francisco
When I’m not under high pressure I try to wake up early and spend at least one hour before work exercising.
Because I’m a workaholic I can’t do it without panicking that I’m not working
So I found a trick: I consider a morning routine as a 1-1 with the most important employee — myself.
Life insurance companies buy garbage data center bonds mixed with high quality assets keeping them A+ rated and life insurance companies protected by USA government they will buy back defaults
This is not 2008, this is 2026
Enjoy and live with that
With Andrew Ng on AI Fund event
Читать полностью…
Creative earnings are +260% more for AI Adopters vs non-adopters
Читать полностью…
Kevin invited me to try HYROX, a fitness face we did nearby Golden Gate Bridge. 6am, Sunrise. So beautiful.
Читать полностью…
Got 97 views and 1 like on my last tweet. But the like is from Ed Chi, VP Research DeepMind
Читать полностью…
Today, ex-openAI Diogo Almeida introduced Jev — a new class of ML models.
He claims it’s the jump from LLMs, same as LLMs did to RNNs (pre 2017)
The idea is to predict next category probability, instead of predicting next token.
Example:
Question: “I paid twice, which email I should use?”
Variants: billing / technical / returns.
Result:
billing → 0.97
technical → 0.01
returns → 0.02
Why it is fast?
All probabilities can be calculated in parallel, it is much faster then sequential token-by-token in LLMs.
How fast? ~75x faster response time VS GPT-5.6 Terra
How cheap? ~170x cheaper vs GPT-5.6 Terra
You can watch a 4 min video, it’s wow!
What are my thoughts about it:
It’s impressive: for gaming, robotics, data scrapping, any environment where number of posible actions is limited that is huge: for example a real-intelligence video game will be only $7/hour in tokens (they messured it on Doom)
What is a TL;DR about it? I would make a joke and say it’s a Categorized Markov Chain trained on large corpus beating best LLMs on cost and latency for structured decision tasks
Really-really impressive idea and implementation!
If you’re interested in math, it goes from 100% theoretical science to a science with experiments. https://www.youtube.com/watch?v=rB9YOi3lb7w
Читать полностью…
SF is the best place for building ❤️
Читать полностью…
My friend Sasha Malysheva has built a great team, and they do intermodel communication on the latent space level. Worth reading!
https://www.wired.com/story/russian-startup-mostik-ai-models-communication/
Random reflections on mathematics.
In my youth, I would make jokes claiming that 2+2=3 or 2+2=5. My teachers disapproved, and I was, of course, fundamentally incorrect.
We begin with the numbers: 0, 1, 2, 3.
This is how we learned to define quantities, such as stating that there are five people in a group or that we have killed one mammoth
We then learn addition: 1+1 = 2, 1+1+1 = 3. Occasionally, we learn the inverse operation, such as 1 = 2 + (-1). This is how we learned to count
For instance, if one more person joins us, we now have 5+1 = 6 people attempting to hunt a mammoth.
Next, we learn multiplication: 1 * 2 = 2, 2 * 2 = 4; and sometimes division: 6/2 = 3, along with the inverse operation: 1/2 is the inverse of 2. For example, one mammoth is sufficient to feed 16 people, meaning one person can consume 1/16 of the mammoth.
Modern mathematics studies these concepts as follows:
• Formal Arithmetic: We have 0 and 1, and the operation of increment (+1).
• Group Theory: Studies the operation (+) and certain constraints (e.g., a+b = b+a).
• Field Theory: Studies the operation * and sometimes /.
• Field Extensions: Studies the addition of new elements, such as i, where i * i = -1. Sometimes, this can be expressed as ℂ = ℚ(√-1), yielding the well-known complex numbers, or more advanced structures like quaternions ℍ, where i * i = j * j = k * k = i * j * k = -1.
This progression leads to: ℤ⁺ ℤ ℚ ℂ ℍ …
Studying field theory provides a highly abstract concept: one should never think about numbers, but only about operations. For example, one can define a seemingly unusual multiplication operation as i * i = -1, which makes no sense in the context of hunting mammoths. Similarly, quaternions are very abstract, yet they remain essential for physics.
The next key point is as follows:
1. We have one of the great inventions of the 20th century: the calculator. It is excellent for performing +, -, /, and * operations on numbers.
2. Now, we have AI. AI is excellent for performing +, -, /, and * operations without specifying what +, -, /, and * mean.
For example, on elliptic curves, there is a sum of points, where you take a point and find the next point on the elliptic curve, creating the sequence: P, P + P, 3P, 4P … You then take another point Q and create the sequence: Q, 2Q, 3Q. This addition operation is defined by the tangent to the point on this elliptic curve.
To the surprise of many, this task requires computation - specifically, knowledge of how to construct the tangent and how to find the next point. It is noteworthy that our generation often cannot even multiply mentally, let alone handle elliptic curves used in the most advanced cryptography.
Therefore, despite the extreme complexity and abstractness of mathematics, particularly addition on elliptic curves, there is also multiplication on them, which is even more complex for humans.
However, today, in an era characterized by:
1. Computers that allow for the rapid verification of whether addition on an elliptic curve is correct
2. AI that can perform formal reasoning
We obtain mega-calculators for tasks such as computing minimal elliptic curves with maximal rank, as mentioned above.
Do you understand the implication? We have formalized arithmetic, addition, and multiplication, moved away from digits, and created a mega-calculator that is also capable of formally deriving proofs.
In other words, AI has not invented anything new; rather, we have begun solving a new class of problems—more abstract, but still using the same method: the calculator of the 21st century.
And depends on the task: count mammoths, or lock a super secure safe locker 🔒, you have a sum function that the calculator can use.
And what about 2 + 2 = 3, and 2 + 2 = 5.
I was wrong. In standart arifmetics it is 2+2 = 4, but you can also consider Z_3, where all arifmetics is by module 3, and write 2+2=1
AI commoditizes the execution of thought, but not the direction of thought.
Intelligence used to be the most important human selection factor; our IQ grew over the last centuries. However, with access to the best models, we have a pocket Einstein available for almost free. Thus, the idea is that intelligence is indeed cheaper; AI has only changed the price for the execution of intelligence. Meanwhile, the cost of choosing the direction has gone up
it is getting more competitive to define first principles, taste, judgment, and agency.
creation is the cheapest today.
choice is the most expensive.
• First principles (axioms) define what is fundamentally true about the world
• Taste tells you what is good and what is bad
• Judgment helps you choose what to do when the answer is uncertain
• Agency is your ability to turn decisions into reality
A small podcast with me
https://youtu.be/S4m9c8HivQQ?is=r_nEhsYhMesfnf6R
Linear vs Exponential.
Old meta:
I work 6 hours I get $100
I work 12 hours I get $200
New meta:
I work 6 hours I get $10
I work 7 hours I get $1M
why?
We used to think intelligence can scale linearly, simply hiring more people, working more hours.
now you can scale Einstein level people, by deploying AI models
how do we observe it today?
• Work as a SWE? There is a risk even a $200 subscription can be better than you
• Work at Anthropic/OpenAI? Your salary is $300k - 2M
leverage.
I find Seedream Pro best for realistic, and MAI 2.5 for higher quality portraits
while Nano Banana has a very well-known style
I rewrote 90% of the app in 6 weeks solo.
Now varg.ai is an OS for creative teams:
Create an API key. Hand it to Claude. Video & images land in your team's storage
"Copy for Agent" → reuse in code or web app, by you or another agent
Creative people and AI agents, finally in the same space
Everyone discuss today that Anthropic bulk-buying rare books and destroy them.
For example, ISBNdb's website literally says "'AI company destroys two million books'
When a scandal started, Elon Musk immediately reacted: “I’ve asked the SpaceXAI team to preserve any rare books in a library and scan them the hard way vs just cutting off the spine and scanning”