Boring Computer Use for legacy systems | CEO of @IlmenAI (Techstars ’24, Xoogler, ex-Microsoft) EB1-A Gateway between old systems and Super Intelligence.
San Mateo, CA
Joined May 2008
- Tweets13.7K
- Following3.4K
- Followers1.9K
- Likes39.4K
Daniel Kornev retweeted
Introducing Flint, @Recursive_SI’s first model: a fast local coding model built with our automated research system, the Eureka Machine.
• Best-in-class fast local coding
• Up to 2x faster than comparable models: 125+ tokens/sec on a MacBook Pro, 150+ on RTX Spark
• Built from scratch in under 4 months
Private, unmetered intelligence on your own machine. Open weights available now on @huggingface, run it locally on your favorite framework or try it for free on @opencode or @pidotdev.
Daniel Kornev retweeted
Windows 11 Copilot+ PCs are set to gain new on-device AI capabilities in the coming months, as announced at Microsoft's Windows event a couple weeks back.
windowscentral.com/microsoft…
Exclusive to Copilot+ PCs, Copilot will be able to understand context, take action, and lean on local AI models without always reaching out to the cloud, though it will still be able to do so when necessary.
I think this is Microsoft's natural evolution of Copilot+ PCs. It's been a while since Microsoft last built any experiences that lean on the NPU, and it sounds like Hybrid Intelligence inside Copilot will be the first set of AI features to do this (at least in part) in over a year.
We already know that Microsoft is slowly moving away from Copilot+ PC as a marketing brand, and that makes me wonder if "Hybrid intelligence" is the new "AI PC" identity that Microsoft wants to get in front of buyers looking for their next PC.
Daniel Kornev retweeted
Hardware Partner Spotlight: NVIDIA joins The @Google Gemma Hardware Hackathon @ AGI House!
Build with Gemma. Run it locally on @NVIDIARobotics Jetson. Bring your next edge or physical AI idea to life.
On October 24, join 100 selected builders, the @GoogleDeepMind Gemma team, and hardware partners for a full day of hands-on building—from camera assistants and intelligent edge devices to robotics applications.
🤖 NVIDIA hardware available to borrow during the hackathon:
• 15 Jetson Orin Nano Super developer kits, each with an NVMe SSD
• 15 Jetson Thor developer kits
These kits are available for on-site development and must be returned after the event. Availability is limited.
🏆NVIDIA hardware awards are also in the works—stay tuned for the final prize details.
Every project must use a Gemma-based model and run entirely on-device. Bring your laptop, your ideas, and your robot if you have one.
Compete for a share of $45,000 in cash prizes, plus partner hardware awards.
📅 Saturday, October 24, 2026 · 10 AM–5 PM PT
📍AGI House · Hillsborough, CA
Hosted by Google DeepMind and AGI House
Apply to build → agihouse.org/events/the-goog…
#Gemma #NVIDIA #Jetson #EdgeAI #PhysicalAI #AGIHouse
Daniel Kornev retweeted
Patrolling the Mojave isn't the only thing making us wish today. Happy 16th birthday, #Fallout: New Vegas!
Join us, @Obsidian and Victor in celebrating Couriers who just won't stay buried 🤠
![An image from Fallout: New Vegas showing dialogue with Victor wearing a party hat outside the Prospector Saloon. Dialogue options read:
Happy 16th birthday, Victor!
[Old World Gourmet] Stuff face with birthday cake
[Wild Wasteland] Roughly how many times have you dug me out of the grave by now?](https://pbs.twimg.com/media/HU_64NzW0AAyh4r.jpg?name=orig)
ALT An image from Fallout: New Vegas showing dialogue with Victor wearing a party hat outside the Prospector Saloon. Dialogue options read: Happy 16th birthday, Victor! [Old World Gourmet] Stuff face with birthday cake [Wild Wasteland] Roughly how many times have you dug me out of the grave by now?
Daniel Kornev retweeted
new paper from USC, UC Irvine and Intel Labs on reusing a small model's prefill inside a big one: the small model reads the prompt, a linear map translates its KV cache into the big model's format, and the big model recomputes only part of the positions
Qwen3-0.6B prefilling for Qwen3-14B (23x apart in size) with 30% of positions recomputed: GSM8K goes from 71.6% with the map alone to 90.2%, against 95.1% for the 14B on its own, and prefill runs up to 3.04x faster
the map alone can leave the big model worse off than the small one: Llama 3.1 8B into Llama 3.3 70B scores 81.0%, below the 8B by itself (82.7%), and recomputing 40% of positions brings it to 92.0%
section 2.2 is worth reading on its own: choosing which positions to recompute by attention score does worse than choosing at random, because the 14B puts 72% of its attention on the start token and 24% on the chat template and instruction, which leaves 4% for the actual question (at a 20% budget the attention-based selector doesn't pick a single question token)
so attention weight mostly tells you where the sink and the template are
arxiv.org/abs/2610.11358
Daniel Kornev retweeted
You can’t automate what you can’t measure. This means that evals are one of the gates to diffusion of AI in the enterprise.
We can test our deterministic processes through software, but most enterprises have no useful way of understanding how their non-deterministic processes are working today. Specifically the work that agents are doing for them.
Evals are mission critical for enterprises adopting AI because you have no other way of knowing what’s working, what’s broken, what changed, what improved, what you can do more of, etc. if you don’t have a good sense of how agents work in your environment today. All changes, upgrades, and deployments are downstream from good evals.
Not only are we going to get vastly more domain specific evals over time for the labs and across the industry, but every enterprise will also need a clear sense of how agents are performing in their environment as well. Huge opportunity.
Just spoke to one of the big data labeling businesses.
Few interesting insights:
- They predict the majority of their revenue will come from Fortune 1000 enterprises, not labs in a few years
- They believe every company will want to own their intelligence, but owning intelligence does not necessarily mean using open source models
- A company’s evals will become their main proprietary IP given the improvement in agent performance after properly setting up & running internal eval environments
- Most enterprises haven’t graduated from coding agents and it’s largely due to not having the proper eval infrastructure to make non-Eng agents performant
Daniel Kornev retweeted
2022: the ChatGPT shock
Generative AI
LLM
Hallucination
Chatbot
Prompt engineering
Diffusion / Stable Diffusion / Midjourney
RLHF
2023: the tooling and fear year
GPT-4
Multimodal
Foundation model
Fine-tuning
Open-source LLMs (Llama)
Plugins and Custom GPTs
Copilot
Embeddings and vector database
RAG
Context window
Jailbreak
AI alignment / AI safety / AGI
Deepfake
2024: from chat to reasoning
RAG (peak hype)
Function calling / tool use
Small language models
Long context (1M tokens)
Text-to-video (Sora)
Chain-of-thought
Reasoning models (o1)
Test-time compute
Synthetic data
AI slop
Computer use
MCP (Model Context Protocol, launched in November)
Evals
2025: the agent year
Agentic AI / AI agents
MCP going mainstream
Vibe coding
DeepSeek moment and distillation
Context engineering
Deep research
Agentic coding (Claude Code and similar)
Sub-agents and multi-agent systems
Agent skills
A2A
Agent memory
Sycophancy
Spec-driven development
World models
AI bubble
2026: harness and beyond
Agent harness / harness engineering
System One models / Jev
the buzzwords moved from the model itself (LLM, hallucination), to how you talk to it (prompt engineering), to what you feed it (RAG, context engineering), to what it does (agents, MCP), to the system around it (harness).
Jev now questions whether every decision needs a chat-style LLM at all.
Daniel Kornev retweeted
Check out the preliminary program itinerary for the 2026 International Mars Society Convention (Oct. 22-24) at @USC. Three packed days of Mars science, exploration, technology & the future of human settlement. Start mapping out your convention today! 🔴 convention.marssociety.org
Daniel Kornev retweeted
🔥 Introducing a Jev Primer & Playground
If you want to learn how Jev works, look no further.
This is the only interactive guide you need to get started.
No more confusion about what Jev can and cannot do. Plus we built a custom Jev Playground for you try out use cases now.
Daniel Kornev retweeted
Learn to build a harness, folks.
It's not surprising to me that so many YC builders want to build domain-specific harnesses.
If you work long enough on a domain-specific problem, you quickly realize the opportunity. But you also realize how important that harness will be to stay competitive in the agentic era.
From a product perspective, harnesses open up interesting new surface areas and experiences for the services/products you provide.
From a technical perspective, harnesses are how you build and maintain a framework and set of best practices for how your users/customers interact with what you offer.
Understanding how to build and design a harness means you can build much stronger intelligence stacks, given that you can customize it and understand the domain well. That's extremely valuable. It may not seem apparent yet, but a harness wave is coming.
If you are getting started, give this list of harness papers to your agents and start upskilling: academy.dair.ai/papers/colle…
If you are already a builder, try building one for your specific domain. It's a lot of fun, and you learn a lot of interesting things to enhance your current agentic tools.
Daniel Kornev retweeted
This is going to be one of my favourite early-stage investments 😊
I met Sasha Malysheva in San Francisco a couple of years ago – through mutual friends at OpenAI. A couple of hours later, I realised I’d actually seen her before – when we were both kids, attending the same humanities-oriented Gymnasium in St. Petersburg, a few years apart. That alone was already a little unusual. Not many kids there were particularly technical.
Later I discovered that she also went to the same deeply technical high school I would occasionally attend, and then studied at the same university as I did – again, a couple of years later.
She eventually ended up at JetBrains, which for years was everyone’s favourite place for programming meetups in St. Petersburg. Whenever I came back from my travels, one of the first things I’d check was whether JetBrains was hosting anything interesting. You’d think the coincidences would stop there, but somehow she also ended up calling Armenia home for a while… At some point, the number of coincidences became statistically suspicious 😅
But the much better reason to invest is what Sasha and the Mostik AI team are actually working on.
Their research direction around latent representations is especially interesting to me given our own work on Unum's UForm models. So when the opportunity came to back them through AAL, there really wasn’t much deliberation involved.
Cherry on top: they’ve just taken first place on ARC-AGI-3 leaderboard, they have a Fields Medalist as a Chief Scientist, and possibly one of the cutest names among the emerging Neolabs – mostik.ai ☺️
Very happy to be backing Sasha and the Mostik team through @aal_vc, together with @stepa and @EugeneMkhlv!
meet @mostik_ai!
what happens when you put 12 PhDs in one room for four months? first place on the ARC-AGI leaderboard, which I can't say much about while the competition is still running. and this, which I can.
everyone's arguing about whether open models will catch up to frontier models. we think it's the wrong question. here's the one we pose: why does a frontier model have to generate your answer at all, when the only thing you need from it is the reasoning?
we do this by enabling models to communicate in latent space. through our protocol, hidden states pass straight from a frontier model into a small one running on your infrastructure -- no text between them, and neither model is fine-tuned. two models from different families, sharing reasoning, both left untouched.
how do we know it works? we tested it on a setup where a 753B model reads the problem, and a 4B edge-class model writes the answer. with this approach, we get results 80% as accurate as the frontier model, but at 20x faster performance.
we're committed to preventing frontier model lock-in and are already partnering with inference providers to accelerate open-weight adoption. we've done this between 15 of us, in four months, 12 PhDs and a Fields medalist, backed by @generalcatalyst
WIRED has the first external account of the company and the work: wired.com/story/russian-star…
full writeup, the setup, and all the numbers: mostik.ai/read-more
Daniel Kornev retweeted
Everyone's suddenly building company brains.
Nobody agrees on what's inside one. 😵💫
So we opened up 9+ company brains to see how they're actually built.
Every single one does the same four things: getting signals, remembering, dreaming & pruning, speaking & searching.
𝟭. 𝗚𝗕𝗿𝗮𝗶𝗻
Garry Tan's open-source personal brain. Your email and calendar flow into a git repo, and a nightly job re-links everything and flags what's gone stale.
𝟮. 𝗺𝗲𝗺𝟬
A memory library you call from your own code. It only stores what you explicitly tell it to, and ranks fresh facts above idle ones at search time.
𝟯. 𝗟𝗲𝘁𝘁𝗮
For building agents that remember across sessions. The agent decides what's worth keeping, and a second agent tidies up its memory in the background.
𝟰. 𝗭𝗲𝗽 / 𝗚𝗿𝗮𝗽𝗵𝗶𝘁𝗶
A knowledge graph with a clock in it. When a fact changes, the old one gets an end date instead of being overwritten, so you can still ask what was true last March.
𝟱. 𝗦𝘆𝗹𝗽𝗵
A content brain that lives entirely in a git repo. Agents write drafts, humans publish, and afterwards the agent reads your edits to learn what it got wrong.
𝟲. 𝗗𝗜𝗬 (𝗖𝗹𝗮𝘂𝗱𝗲 𝗖𝗼𝗱𝗲 + 𝗴𝗶𝘁)
What most engineering teams actually do. Markdown in the repo, grep instead of search, and pull requests as the only thing keeping it honest.
𝟳. 𝗣𝗹𝗲𝘁𝗼𝗿
A brand brain for marketing teams. Campaigns, assets and performance data in one tree, and the brand rules only move when a human signs off.
𝟴. 𝗚𝗼𝗿𝗴𝗶𝗮𝘀 𝗖𝗼𝗿𝘁𝗲𝘅
Built in-house by an eight-person AI team. 12,000 markdown nodes in GitHub, and every night the questions it got wrong become PRs that fix it.
𝟵. 𝗦𝗹𝗶𝘁𝗲 𝗔𝗴𝗲𝗻𝘁
For teams whose knowledge lives in docs and across sources. It watches ~20 connected tools (Slack, Drive, GitHub, Jira, etc.) for what's gone stale and sends the diff to whoever owns the page. Nothing changes without human approval.
We just launched an interactive ebook with architecture notes from real 149 teams of builders and users, interview insights, and the complete research.
The ebook is free, get it here: slite.com/ebooks/company-bra…
Which brain would you pick? 🧠
Daniel Kornev retweeted
new paper describing BDH-CQ from the Pathway team: a 150M model that gets 29.5% on ARC-AGI-1 at $0.0007 per task (less than a tenth of a cent!) -- it's a new best on cost efficiency
demonstrations write into a recurrent memory, then it iterates in latent space and decodes only the answer, and no intermediate reasoning ever gets decoded into text
I'm wondering how much of this carries over to ARC-AGI-3, which is running right now and works completely differently, because there nothing is demonstrated to the model, instead it should go find out the rules by acting in the environment
their memory update might not care whether what arrives is a demonstration or something you discovered by poking around, so it seems promising, but their weak spots are ordering and composition, and the model often solves one test input of a task and misses the others, and a long interactive rollout leans on all of that
arxiv.org/abs/2608.09888
Daniel Kornev retweeted
Very interesting new work from Microsoft.
(bookmark it)
This work is related to this emerging theme of leveraging harnesses for model post-training.
Modern agents run inside a harness that owns tools, context, and control flow. When you train them, the harness owns the environment loop and the trainer only sees LLM request and response pairs.
How it works.
Agent Lightning v1.0 connects any harness to RL through an endpoint proxy in about 3,500 lines, then works through what breaks in that setup, retokenization, sample merging, advantage calculation, loss normalization, and backend scheduling.
Using 6K training examples and modest compute, it moves Qwen3.5-9B on SWE-bench Verified from 41.8% to 56.4%.
Paper: arxiv.org/abs/2608.17528
Track more trending AI papers in our academy: academy.dair.ai/
Daniel Kornev retweeted
No spacesuit required. Just curiosity. Join us this October at @USC! marssociety.org #mars2026 #globalconference #losangeles #science #marsexploration #research #themarssociety #humanstomars
Daniel Kornev retweeted
It makes sense to optimize model routing at the harness layer instead of the gateway layer if you want to hillclimb on accuracy/cost for any e2e task.
Every task is solved by a combination of a model mixture and agent harness. Every task requires a different mixture of models (+harness logic) to be at the pareto frontier of accuracy and cost.
* If you only optimize the model mixture at the gateway layer, you lose the broader context encoded in the harness and only optimize at the LLM completion layer.
* If you optimize the model mixture at the harness layer, you can make the "optimal" model choices a priori and while in the agent loop
For any given task, the model and harness are probably co-optimized together, so that the model mixture can only exist with this specific harness shape and vice versa
Daniel Kornev retweeted
LLMs made meaning measurable: concepts became vectors, so how close two ideas are is a number you compute now rather than of a thing you argue about :)
the obvious next question is whether different models agree on those numbers: the Platonic Representation Hypothesis said they do (models converge toward a shared representation as they scale), however that got challenged this year -- Gröger, Wen and Brbić at EPFL (ICML 2026, arxiv.org/abs/2602.14486) showed most of the reported convergence was width and depth mechanically inflating the similarity scores, and correcting for that makes the global trend disappear
but they also showed that models still agree on which things sit near which other things, across image and text and video, trained by different labs on different data (the Aristotelian Representation Hypothesis)
and local agreement is already enough to translate: vec2vec (Jha, Zhang, @shmatikov, @jxmnop at Cornell) moved resulting embeddings between models that never trained together, and got 0.92 cosine similarity
embeddings-at-inference-stage are the narrower case tho: the general version is one model handing another something richer than a sentence, which is a much harder problem than the one vec2vec solved
whatever ends up doing that translation sets the terms for how everything else connects, but I'm sure it will be like a Rosetta Stone for machine minds 🔥
in ten years I don't think frontier networks will look anything like transformers
right now every model is one block copied a few hundred times: identical layers in a uniform stack, and that's a reasonable engineering choice, because homogeneous things parallelize cleanly and are trained conveniently
but "natural" intelligence doesn't work like that: cortex and cerebellum are radically different architectures running at different speeds with different memory behavior, and there are hundreds of distinct neuron types spread across the whole thing
my bet is that frontier systems within a decade will be deeply heterogeneous: e.g. a memory network with effectively unbounded context sitting next to a fast reflex network that answers in milliseconds, and they both trained jointly as one system
however, modules with different shapes and different timescales make gradients behave badly, and there's no good theory yet for training that whole thing at once
I think the path runs through the geometry of latent spaces, which is roughly what I've spent this year reading about: if modules with different architectures have to train as one system, you need to know what the space they communicate through actually looks like
homogeneity was a concession to the hardware and the optimizers we had at the time, and currently they both keep changing
Daniel Kornev retweeted
This paper is f*cking genius
A computer science paper from Stanford AI Systems Lab formalizes why multi-agent pipelines fail and introduces the Control Plane Pattern for autonomous research
The result: cuts investigation latency from 4 weeks down to 30 minutes while reducing token usage by ~10x with zero context handoff decay
The crazy part is how naive multi-agent chains degrade context at every handoff - causing Synthesis agents to recommend field visits for a payer tier problem
Deterministic signal queues, single centralized reasoning agents, and Knowledge Graph control planes replace brittle handoff pipelines
Most AI teams build multi-agent pipelines where specialized agents pass context down a broken chain
This paper structures agentic analytics into a bounded, graph-constrained operating system
Read the complete paper + article below
Bookmark it for future reference
Daniel Kornev retweeted
AI has had exactly two scaling axes that worked so far, and the second one is starting to look finite too
the first one was pretraining: with scaling parameters and data, we got world knowledge (i.e. ChatGPT had read enough to know things), but it started saturating a while ago
the second one was RL, and people had been doing RL the whole time before that: RLHF is RL but it never scaled far because it was trying to control the exact output, which tokens come out, how the text reads, but you can only push that so far before you’re just polishing
RLVR dropped that constraint: giving the model a task, then checking whether the final answer is right, and ignoring everything in between -- so the model does whatever it wants in the middle and only the endpoint gets graded, and that’s much closer to actual RL and it’s what bought us planning and reasoning
(arguably, tool use sits around 2.5 on this list -- while useful, it's not a different kind of thing)
so one axis gave knowledge, the other gave reasoning, and both of them are one model working alone
the next axis is how many models you can get working on the same problem, which is a different kind of axis than the previous two
we know that multi-agent RL has always been the harder problem: I spent years in that literature and the gap between single-agent and multi-agent is definitely not incremental -- it’s a whole different class of difficulty! which is also why the derivatives are steep at the start, nobody has picked the easy wins yet...
and the thing that gates this multi-agent coordination is communication: models can only coordinate as well as they can exchange information, and right now they do that by writing sentences to each other
imagine what could we possibly achieve if we properly open that third axis development by letting models to exchange information in their native "language" without loosing any computational data that they produce during inference
on Thursday, @taalas_inc agreed to join AMD
they etch a model's weights directly into the transistors, so there's no external memory to read during a forward pass and the memory wall that caps every GPU inference system doesn't apply: their demo chip runs Llama 3.1 8B at ~17k tokens/sec, where an H100 does about 150
the objection people raised back in February was that a chip like this only runs the model it was built for
but the weights go into mask ROM permanently while the KV cache and LoRA adapters live in SRAM where they can still change -- so there's already a frozen half and a moving half
right now the frozen half is an entire model, and when the model changes they respin the metal layers, roughly two months of turnaround, and they've priced three of those revisions into a four-year lifespan
nobody's tried making the frozen half smaller than that: you'd need some way to connect the parts that stay put with the ones that change, which is its own hard problem
but if that existed you'd only re-etch whatever actually moved, and a chip could last past one model generation
and all of this is seven months after Nvidia paid $20B for Groq