Did you know that Toyota has a footprint in modern GenAI frontier?
Toyota Research Institute is one of the folks who is pushing the boundaries of Large Behavioral Models. Yeap, you have not missheard it. LBM - large models that are very similar to LLMs. They are trained on motions in 3D space if I put it simply.
You can fine tune those models and build your own robots that focus on Kitchen tasks, the ones that are specialized in yard tasks ect.
That is one of the reasons why they say that 2026 is Robotics era kick off .

https://www.tri.global/our-work/large-behavior-models

How I accedentally discovered Super Power of Comet browser.

I wrote an ADR(Architecture Decision Record) with Windsurf Claude Sonnet to change the feature we already have. I got a bit lazy and wanted somebody to read the ADR outloud.
I opened numeo Github Repo on my belowed browser Comet to check if it could read the MD file outload for me. Comet suprized me immediately by identifying the PR page, and failing CI steps. That triggered me to think... "Could I ask Comet browser to review our PRs??

So I tried.... It started going through all the PRs one... by... one....

Topic: AI Agent reliability #agentreliability

MAKER, developed by Cognizant AI Lab with UT Austin, is an agentic system that completed a task with over one million steps and zero errors, addressing the reliability limits of current LLMs for long, chained reasoning. Unlike the usual focus on increasing model intelligence, MAKER uses extreme task decomposition: each tiny subtask is given to a dedicated microagent, and agent outputs are error-corrected by local multi-agent voting and red-flag filtering. This modular framework makes reliability scale logarithmically with the task length (Θ(s log s)), instead of suffering exponential error growth, enabling small and efficient models to outperform larger ones on extended tasks. The success of MAKER suggests that high reliability in large-scale AI can come from orchestration and structure, not just bigger models

Link: https://arxiv.org/abs/2511.09030v1

One of the most practical book to bulid RAG I have encountered lately. It shows challenges around RAG and how to address them step by step. One by one.

https://www.youtube.com/watch?v=-Tgc_9uYJLI
Thisis a big leap forward in On-Device / Edge infrence. This model can work directly on your devices or browser.

Our Agents @Numeo have to have lots of skills/instructions. This comes with its challanges. You probably have heard about "Curs of Instructions" paper that explains why having too much instructions is bad.

At Numeo we have been doubling on building our Agents in a way so they can "Pick Up" skills/instructions progressively. We will defenitly experiment with Function Gemma.
I am expecting our tool calls will be much more precise, much faster and much cheapper with this model.

We are living in a world where
- Resumes are generated by GenAI,
- Applications are submitted by AI agents
- Replies are written by AI agents :)

STEP-3.5-FLASH: Breaking the Speed vs. Intelligence Trade-off

StepFun released a model with 196 billion parameters, but only 11 billion activate per token. This is a sparse Mixture-of-Experts model. The model is performed pretty well on SWE-Bench. People are claiming that to be GPT-4 level quality which is pretty good based on my experience.

Its throughput is 300 tokens per second on modern hardware. Compare that to DeepSeek's 33 tok/s. You're looking at roughly 9x faster inference on the same task, with higher accuracy on coding problems.

Context window: 256K which is pretty cool. Handled through a 3:1 sliding window attention mechanism that avoids the quadratic complexity that kills most long-context models.

Why This Matters for Agents

The model uses multi-token prediction—predicting three tokens per forward pass instead of one. Sounds simple. It's not. This breaks the serialization bottleneck in autoregressive generation and fundamentally changes the latency profile for agentic workflows.
For autonomous agents, this matters. You need models that reason deeply but respond quickly. Step-3.5-Flash handles both without the usual trade-off.

Local Deployment

Runs on a Mac Studio M4 Max with 128GB. You get GPT-4 class reasoning entirely offline. That's a shift for privacy-sensitive work, edge deployment, or cost-critical inference.
INT4 quantization is available. Cuts memory further without catastrophic loss in capability.

Real Benchmarks Beyond Coding
🔗 GitHub: https://github.com/stepfun-ai/Step-3.5-Flash
🔗 Hugging Face: https://huggingface.co/stepfun-ai/Step-3.5-Flash
🔗 NVIDIA NIM: https://build.nvidia.com/stepfun-ai/step-3.5-flash
🔗 OpenRouter API: https://openrouter.ai/stepfun/step-3.5-flash
🔗 Original Blog: https://static.stepfun.com/blog/step-3.5-flash/

We are hiring from Uzbekistan

My laziness hit a new rock bottom :).

Now Claude Code is reading its output out loud:).
So I don't have to read small text out of my screen all the time :).

Local TTS
https://github.com/ktaletsk/claude-code-tts#

So true. There is not much MOAT in software anymore.

I bet we are gonna see lots of Software companies start adding Hardware products to create a defensibility very soon.

AGI is here folks! Holly cow!

Claude Opus 4.7 did it!

https://www.youtube.com/watch?v=kYkIdXwW2AE
If you are aware, Yann LeCoun has left Meta and actively researching/developing and investing in other startups that are trying to move frontier in NON-LLM directions. Yann isinvesting / betting on World Models (Generic name).
He thinks LLMs are not the best proxies of the way Human intelligence works.

On this video, Yann is sharing his thoughts on JEPE models. Problems of Video generation approaches. Why Autoregressive models are not great for Video gen cases.

Enjoy folks.

https://www.youtube.com/watch?v=xOP1PM8fwnk

One of the best explanations of how current Video Gen models are built.
This is very interesting space to be in. You learn how H264\265 codecs are built. How to leverage the hybrid of Autoregressive and Diffusion models to build a time series of images. (Video is series of images over time + Audio).

The video quickly touches FFT (Fast Fourie Transform) as well. FFT is a foundation of signal processing. The most of the compression algorithms like JPEG, MP3 are build literally on top of Fast Fourie Transform.

I ihghly recommend watching it.

https://www.linkedin.com/posts/bnhop_we-run-our-coval-stt-benchmark-every-30-minutes-ugcPost-7483956231968018432-cVqB/?utm_source=share&utm_medium=member_desktop&rcm=ACoAAAzP3TIB4pvnEUAvi_ABR6D5qCr_LxAIk0s

Btw folks, follow her to get interesting Voice AI related stats weekly. She is a Ex-Tesla, and currently building a coval.ai, voice ai eval solution.

The internet has accumulated a lot of knowledge in the form of PDFs. Some of these are digital-native PDFs, while others are just wrappers around images. Until now, there hasn't been much demand for parsing these PDFs in a structured way.

Document parsing has become much more popular in recent years. We wanted to leverage all that accumulated data to train our LLMs. This fueled interest in building OCR systems that understand tables, images, and mathematical formulas well.

At Numeo, we process around 50,000 freight documents (Rate Cons, BOLs, etc.) daily. I've been researching how to improve our document processing pipeline, and my research results surprised me quite a bit.

It turns out all top 5 OCR solutions/models come from China, according to the OmniDocBench (https://github.com/opendatalab/OmniDocBench).

China is cooking...

🚀 Hiring — Software Engineer

Numeo AI is redefining logistics in the US. We're building tech that powers supply chains at scale. Growing fast—really fast. Looking for people who thrive in high-intensity environments.

Who we're looking for:
• Deep technical expertise (React Native, React, Node.js)
• Builder's mindset—wear multiple hats, move fast
• Thrive in high-intensity startup environment
• Hungry to grow rapidly
• Work in our Tashkent office—in person

What you'll work on:
• Build and scale mobile apps (React Native)
• Architect backend systems (Node.js)
• Develop frontend interfaces (React)
• Own full-stack features from conception to production
• Ship features that solve real logistics problems

Must haves:
• Proven experience with React Native, React, or Node.js
• Full-stack capabilities
• Strong software engineering fundamentals
• Experience shipping in fast-paced environments
• Passion for logistics problems

The deal:
• 10–12 hours/day—we're building something massive
• On-site in Tashkent office
• Competitive compensation based on impact
• Rapid growth opportunities as we scale
• Work with ambitious builders

Not for job-seekers. For engineers who want to say: "I'm building at Numeo AI."

📩 Apply: jobs@numeo.ai

@paiziev24

Hiring haqida kuzatuvlar

Oxirgi oy ko’p vaqtim Numeo’da hiring-ga ketyapti. Shu jarayonda bir nechta kuzatuvlar bilan bo’lishmoqchiman.

1. AI yozgan CV’lar juda ko‘payib ketdi

Cover letter ham, CV ham AI yordamida yozilgan. Ba’zida CV’dagi achievement-larni nomzodni o’zi ham o‘qimagan bo‘ladi. Interview-da savol bersam bittasi: “Ai yozdi shekili” deb javob berdi😅

2. Chiroyli CVning o‘zi yetarli emas.

Mashhur kompaniya yoki universitet, ko‘p yillik tajriba - albatta zo’r. Lekin yangi narsani tez o‘rganish tajribasi, growth mindset va ownership men uchun ancha muhimroq.

Afsuski, ko‘p nomzodlarda aynan shu tomonlar yetishmaydi.

3. Prezident maktabini bitirgan kuchli nomzodlar ko‘p.

Hozirgacha ulardan ikkitasini jamoamizga ishga oldik. Sifatli ta’lim natijasini amalda ko‘rib hursand bo’ldim.

Menimcha, davlat bunday maktablarga ko‘proq investitsiya qilishi va kam ta’minlangan oilalar farzandlari uchun imkoniyatlarni kengaytirishi kerak.

4. Eng yoqqani - odamlar aynan Numeo-da ishlash uchun kelishyapti.

Ko‘pchilik shunchaki ish qidirmayapti. Ular aynan Numeo jamoasi, founderlar va product bilan ishlashni xohlab kelishmoqda.

Bu esa motivatsiyasi bor odamlarni topishga yordam berayapti. O‘zim ham Numeo’ga aynan shu sabab kelganman.

5. Numeo-da hiring davom etmoqda.

Apply here numeo.ai/careers

+ Ochiq vakansiyalarni bio’dagi linkdan topishingiz mumkin.

@latipovasamira

Hardware vs Firmware vs Software vs Jellyware/Liveware(New terminology)

Quick background
• Hardware is what you can kick. Physical
• Firmware is the software that thinks it's hardware — burned into a device, shipped once, updated reluctantly.
• Software is everything above that: written, reviewed, versioned, blamed on a specific commit author. Every one of those categories assumes a human wrote the source and a human can read it back.

Jellyware is coming, and it will produce the most adaptive software products we've built — because they adjust to different needs differently. We at Numeo, have certain modules that depend on third party services with lots of non commincated changes. Every change on the third party will introduce regressions on our product. We are thinking to convert some part of the product to Jellyware to ensure that part of the code will adjust into changes dynamically.
The logic will be:
• if our logic in that module will product an error, we go back to our Agent to re-crete a version that will work and achieve our goal
• Bring the code back and use it for future interpretations

That that world, what you trade is exactness for flexibility: predictability up to a certain degree, flexibility far beyond anything a fixed binary can offer. Same word, new layer of the stack