Sentius

We build Autonomous Agents for the Enterprise.
Made with 💖 by AGI enthusiasts from US and UK.

[Sentius @SentiusAI]

This is a super exciting technical approach to merchants.

[Mikhail Parakhin @MParakhin]

And now SimGym is available for all the eligible merchants (SNI/Online Storefronts), no waitlist anymore! Sorry it took a while. Happy building!

[Zag Zino @ZagZino]

The stats back it: 90% of startups fail. The ones that succeed don't have smarter founders. They have founders who were still standing after everyone else quit.

Toughness isn't emotional armor. It is the ability to wake up to the same unsolvable problem for three years and not stop moving forward.

The hardest part isn't visible in any pitch deck.

[Paul Graham @paulg]

Someone asked what's the most underappreciated quality in startup founders. I realized I could answer this by asking what's the most underappreciated aspect of startups. That's easy: how hard they are. So the most underappreciated quality in founders is sheer toughness.

[Mikhail Parakhin @MParakhin]

"Auto-Compressing Networks" is such a super-simple idea, yet works remarkably well! It (with ELU nonlinearity) became my standard go-to any time I need a small MLP (embedding transformations, QKV forms, etc.) -everywhere I tried, it was slightly, but measurably, better.

[zak @zak_in_space]

With claude code and agent coding, advantage in software engineering is going from autistic people to ADHD people

[Ishaan Sehgal @ishaansehgal]

Post-YC founders are hitting a paradox nobody talks about.

There's more capital than ever. Accelerators, angels, family offices all writing checks.

But there's also more pressure to figure out what actually works in AI before Google or OpenAI just copy you.

Speed to product-market fit used to be measured in years. Now you've got months, maybe weeks.

You're racing against infinite resources with a finite runway.

The window to prove your moat exists is shrinking every single day.

[Subbarao Kambhampati (కంభంపాటి సుబ్బారావు) @rao2z]

Hmm.. So you can't use LLMs to write code that is not already well represented in the humanity's knowledge closure? Who would've thunk? (Hope this is not a deal breaker in them making Nobel-worthy discoveries by next weekend though..)

[Daniel Kornev @danielko]

💯.

The idea of Full Self-Driving is trivial (‘make the car drive itself’). Execution meant cameras, maps, edge cases → a whole autonomy stack.

That’s how we approach AI Agents too. The idea (‘let AI control software’) is simple. But execution is brutal: UIs weren’t built for machines. We’re mapping them, teaching flows, handling billions of edge cases — FSD for software.

Do that right → every UI becomes a Universal API. Here's the 5-min pitch of how we’re building it at @SentiusAI using first principles:

[Sentius @SentiusAI]

One big love for a nano chat!

(pun intended)

[Andrej Karpathy @karpathy]

Excited to release new repo: nanochat!
(it's among the most unhinged I've written).

Unlike my earlier similar repo nanoGPT which only covered pretraining, nanochat is a minimal, from scratch, full-stack training/inference pipeline of a simple ChatGPT clone in a single, dependency-minimal codebase. You boot up a cloud GPU box, run a single script and in as little as 4 hours later you can talk to your own LLM in a ChatGPT-like web UI.

It weighs ~8,000 lines of imo quite clean code to:

- Train the tokenizer using a new Rust implementation
- Pretrain a Transformer LLM on FineWeb, evaluate CORE score across a number of metrics
- Midtrain on user-assistant conversations from SmolTalk, multiple choice questions, tool use.
- SFT, evaluate the chat model on world knowledge multiple choice (ARC-E/C, MMLU), math (GSM8K), code (HumanEval)
- RL the model optionally on GSM8K with "GRPO"
- Efficient inference the model in an Engine with KV cache, simple prefill/decode, tool use (Python interpreter in a lightweight sandbox), talk to it over CLI or ChatGPT-like WebUI.
- Write a single markdown report card, summarizing and gamifying the whole thing.

Even for as low as ~$100 in cost (~4 hours on an 8XH100 node), you can train a little ChatGPT clone that you can kind of talk to, and which can write stories/poems, answer simple questions. About ~12 hours surpasses GPT-2 CORE metric. As you further scale up towards ~$1000 (~41.6 hours of training), it quickly becomes a lot more coherent and can solve simple math/code problems and take multiple choice tests. E.g. a depth 30 model trained for 24 hours (this is about equal to FLOPs of GPT-3 Small 125M and 1/1000th of GPT-3) gets into 40s on MMLU and 70s on ARC-Easy, 20s on GSM8K, etc.

My goal is to get the full "strong baseline" stack into one cohesive, minimal, readable, hackable, maximally forkable repo. nanochat will be the capstone project of LLM101n (which is still being developed). I think it also has potential to grow into a research harness, or a benchmark, similar to nanoGPT before it. It is by no means finished, tuned or optimized (actually I think there's likely quite a bit of low-hanging fruit), but I think it's at a place where the overall skeleton is ok enough that it can go up on GitHub where all the parts of it can be improved.

Link to repo and a detailed walkthrough of the nanochat speedrun is in the reply.

[Daniel Kornev @danielko]

3/3 If you run a complex industrial application — oil rigs, manufacturing floors, global supply chains — and wonder how to build AI co-workers around it: talk to us at @SentiusAI.

This is our #MasterPlanForAgenticAI.

[Daniel Kornev @danielko]

2/3 Everyone chases Word, Excel, Salesforce. But who automates the long tail — apps with 50K niche users that run trillion-$ industries?

Think Schlumberger Omega in oil & gas. SAP in supply chain. Dassault in aerospace. These tools run the world — but their experts are retiring.

We took a page from @elonmusk & @karpathy on Full-Self Driving: maps come first. Yes, “The map is not the territory.”, but without map, you're wondering. You can’t reliably drive without any sense of a map at all.

You can’t automate without an Atlas.

That’s why we’re building the #AtlasOfIndustrialApps. Real UIs. Real workflows. Real reliability.

“Macrohard for Industry.” Because unlike fake gyms, these apps are hard, and they matter.

[Daniel Kornev @danielko]

1/3 Your next hire might not be human — and it may know Schlumberger Omega better than your best engineer.

AI “co-workers” are moving from hype to reality. Labs like @AnthropicAI and @OpenAI are training models in fake enterprise apps — teaching agents to click through synthetic Salesforce, CRMs, email clients.

The costs are staggering: billions on data, $150–250/hr domain experts, “gyms” for agents.

At @SentiusAI we chose first principles: map and operate against real enterprise software. Failures surface early. Audit trails are real. Reliability beats simulation.

One OpenAI exec put it best: “The entire economy will become an RL machine.”

Question is — will it run on fake apps or real ones?

https://t.co/E5zvUfHagf

[Daniel Kornev @danielko]

Preparing to pitch @SentiusAI at @SelectiveVC in less than 4 min.

Come join us here:
https://t.co/nuEwqTAjlr

(No #Macrohard for now, but we'll talk about our latest breakthrough in #ComputerUse tech for the very first time!)

#AIAgents #ComputerUse #FSDforEnterprise #FirstPrinciples

[Paul Graham @paulg]

People who are good at programming will use AI to take the jobs of those who are mediocre at it. So you should study CS iff you're going to be good at it.

[OpenAI @OpenAI]

Dropping soon.

[Lina Khan @linamkhan]

A great reminder that letting startups grow into independently successful businesses, rather than be bought up by existing giants, can generate enormous value.

A win for employees, investors, innovation, and the public.

[Paul Graham @paulg]

If you want to start a software startup, you should still learn to program. Even if AI writes most of your code, you'll still be in the position of an engineering manager, and to be a good engineering manager you have to be a programmer yourself.

[Jerry Tworek @MillionInt]

This guy gets it

[Andrej Karpathy @karpathy]

I always learn a lot more from in-depth analysis of few random cases over dashboards of aggregate statistics across all cases. Both projections can be helpful but the latter is disproportionately pervasive.

[ThioJoe @thiojoe]

I've done it 😤

A fast Windows Explorer native SVG thumbnail extension written in Rust, with no dependencies besides the official Windows API bindings.

It uses the Windows API to render everything, even with hardware acceleration.

Will publish the repo after more testing.