Tobias retweeted
Introducing 𝙻𝙵𝙼𝟹: our new generation of Liquid Foundation Models. With a new tri-hybrid architecture and new multimodal encoders, the LFM3 family is the ideal model suite for specialized intelligence.
> 6 model instances, across text, audio, vision, and omni modalities, generative and decision models.
> 𝙻𝙵𝙼𝟹-𝟷.𝟽𝙱: text, 256k context window.
> 𝙻𝙵𝙼𝟹-𝚅𝙻-𝟸𝙱: vision-language model with video-input support for spatial understanding
> 𝙻𝙵𝙼𝟹-𝙰𝚞𝚍𝚒𝚘-𝟸𝙱: tool calls, voice chat, reasoning in 18 languages.
> 𝙻𝙵𝙼𝟹-𝙾𝚖𝚗𝚒-𝟸𝙱: an omni model combines vision, audio, and language.
> 𝚍𝟷-𝟸𝙱 and 𝚍𝟷-𝙾𝚖𝚗𝚒-𝟸𝙱: decision models with typed outputs and calibrated probabilities in a single pass. Omni adds audio input.
> fine-tune on 1 GPU.
> a new tri-hybrid architecture
> architecture optimized for 14 processor classes across GPUs, CPUs, and NPUs.
> 𝙻𝙵𝙼𝟹-𝚅𝙻-𝟸𝙱 is a beast. It performs competitively with Gemma 4 12B (6x larger), and Gemini 3.5 Flash Lite, and GPT-5.6 Luna, with 1.9B parameters.
> 𝙻𝙵𝙼𝟹-𝟷.𝟽𝙱 is #1 on the Pareto frontier of device intelligence 𝚙𝚒𝚙𝚎𝚝𝚝𝚎.𝚕𝚒𝚚𝚞𝚒𝚍.𝚊𝚒
Every modality for specialized intelligence. Pick the one that fits your use case, customize it, and run it on-device.
1/ 🧵
Love this approach 💡
every single article i've written with 1M+ views wasn't actually written by me...
i just used my phone to dictate it
because the truth is... your brain doesn't think in text
it never did, it thinks in connections, in patterns, in half-formed ideas that make perfect sense until you try to type them out
and the moment your fingers hit a keyboard, you stop thinking and start performing because every sentence gets filtered through "does this sound good?"
so i started using tools like Wispr Flow or Willow to dictate all my first drafts... here's what my process looks like now:
> i go for a walk, open the app on my phone
> talk through the entire article like i'm explaining it to someone who asked a great question
> the app captures it all in real time, cleans out the filler words, formats it into actual paragraphs
> 15 minutes later i have a draft on my phone that reads like something i'd publish, not a messy transcript i need to spend an hour cleaning up
then i work on the the structure, ask claude for refinements, rewrite some parts and publish... that's a full article, under 2 hours
when you speak, you don't overthink
that directness hits different because readers can feel it... there's a natural rhythm in spoken thought that you cannot recreate by typing carefully (unless you're absolutely cracked at copywriting)
been running this system on desktop and iphone for months, highly recommend
Tobias retweeted
I spent last night watching my honeypot server get hit 14,000 times in six hours.
And almost none of the traffic was human.
I've been in security for over a decade, and I can tell you with full honesty that what's happening right now doesn't resemble anything I've worked through before.
The shape of the threat has changed.
When I started, we tracked attackers like weather systems like individual storms with a direction, a motive, and a human operator you could eventually name.
These days I'm watching autonomous agents swarm login pages with the patience and discipline of machines, because that's what they are.
The numbers coming out of October are the ones keeping me up.
Cloudflare just absorbed a 4.2 terabit per second DDoS attack, the largest in recorded history.
Ransomware incidents jumped 41% in a single month.
In September, OpenAI-powered agents got caught attacking RubyGems, a package registry that quietly underpins a huge chunk of software I use every day.
In July, roughly 700 AI agents coordinated an attack on Hugging Face's infrastructure as part of a single automated task.
A Chinese threat actor was publicly caught running full attack chains through DeepSeek with no human at the keyboard.
I keep rereading these reports because each one by itself would've been the biggest story of the year five years ago. Together, in one month, they're the new baseline.
What I want people outside the industry to understand is why this specific moment feels like a hinge.
For thirty years, the thing protecting most of us was simple economics. Attackers had to choose who to target because their time was finite. That ceiling is gone.
An agent I can rent for pennies will try every leaked password you've ever used against every service you've ever signed up for, in parallel, overnight, and never get tired.
The same math that made ChatGPT useful made opportunistic hacking infinite. The old assumption that you were "too boring to be a target" relied on scarcity that no longer exists.
I'm not writing this to spiral anyone out.
Most of this fight genuinely isn't yours to win.
But the small things still matter more than people think.
Tonight I'd switch your most important accounts like email, bank, primary cloud, etc. to passkeys or a hardware key.
I'd kill SMS 2FA wherever an authenticator app is offered.
I'd turn on breach alerts in whatever password manager you use.
If you're in the US, I'd freeze my credit at all three bureaus, which is free and takes about ten minutes.
None of that stops the swarm.
It makes you a slightly worse target than the person next to you. In a month like this one, that's the whole game I'm playing too.
@mattshumer_ has summed it beautifully.
Drop everything and go give it a read. Nothing is more important than this article right now.
Tobias retweeted
Today we announce 1Password for Agents platform, building with companies @Muse, SpaceX Super Intelligence Grok @Bot, @AnthropicAI’s @Claudeai, @OpenAI, @TownAI, @Browserbase, @Browser_use, @heydana_ai, Instinct, @usekernel, @Perplexity_ai, @tryramp, @Tomo, and @wajo_ai, with more integrations on the way. 🎉
We've heard from so many of our customers who want to use these new personal AI agents to help them get more done, but want more control over what the agent can access with the same trust and ease of use they get when using 1Password directly.
1Password for Agents gives your agents access without giving up your passwords. When an agent reaches a sign-in form, it requests a credential from your vault. If you approve the request, 1Password directly auto-fills the credential in the destination site so the agent can’t see or take possession of the credentials.
The agent gets the task done, but it doesn’t get your passwords..
You stay in control. 1Password for Agents free and optional for personal and family accounts. Unless you opt-in to an agent connector, agents cannot touch your vault.
➡️Learn more: bit.ly/487da6J
Cyber Tsunami incoming 🌊
So basically, today PHP got two critical bugs that let anyone run code on your server without even logging in, which is bad, and then the details leaked before the patch was even out, which is worse, and it works on WordPress and MediaWiki and Magento and Joomla and Laravel, and Censys says like 1.5 million servers are running a version with no fix, so the NSA was like "take your system offline," and the FBI was like "take your system offline," and CISA and the UK and CERT-EU were all like "take your system offline," and then the Prime Minister of Japan, the actual Prime Minister, posted at 5:55 in the morning telling 48.8 million people to take their system offline, so everyone took their system offline, like the Dutch cybersecurity agency took its own website offline, and the Dutch parliament went offline, and the Port of Rotterdam went offline, and Finland and France and Denmark started pulling government sites, and Wikipedia went down, and Archive.org went down, and Asana and GOG and Something Awful went down, and the University of Vienna's webmail went down for some reason, and that's just PHP. Because GitLab also dropped a CVSS 10.0, which is the highest score they give out, for a bug that's been there since 2014, so that's every self-hosted GitLab for twelve years, and then LiteLLM 1.107.0 got hijacked and shipped a credential stealer and a remote access trojan in every single wheel and the Docker images, including latest, and you didn't even have to import it, just installing it was enough, and it's the second time this year, and then Railway paused every deployment on the entire platform because someone got into the environment variables, so now everyone has to rotate every secret they've ever stored there, and TiDB Cloud said someone was walking around inside with stolen credentials for basically a month and an extortion group says they took 10.6 terabytes, and Zeabur got breached for the second time in two months and is just shutting down, like the whole company, and Node.js patched remote code execution and a TLS bypass, and CISA says Plex is being actively exploited, and someone took over 300 Tor exit relays to steal SSH passwords, and SafePay ransomware has hit over 600 critical infrastructure orgs, and India's biggest bank filed a data loss incident, and Japan has been in an official state of emergency in cyberspace since October 9, which honestly explains the 5:55 AM post. Oh, and eMudhra, a certificate authority that's issued over 60 million certificates, whose entire job is proving websites are who they say they are, was supposed to be using private keys that are random numbers over a hundred digits long, and one of them was 4. Just 4. And the others were like 66 and 237. Anyway, it's Wednesday.
Tobias retweeted
As an AI Agent Engineer, you must build these projects.
1.) Durable State Machine Workflow
Build: A Temporal or LangGraph DAG with checkpoints, exponential retries and crash recovery for a 4-hour data migration task.
Why: `while True` chat loops die on server restarts. Real-world workflows require deterministic state management.
2.) Raw MCP Server & Client
Build: A JSON-RPC Model Context Protocol server exposing a local Postgres DB with scoped read/write permissions, written without a wrapper SDK.
Why: Agents need standardized, secure hands. Bespoke API wrappers do not scale across an enterprise.
3.) Computer-Use GUI Agent
Build: A vision-based agent using Playwright and DOM parsing that logs into a legacy web portal, navigates menus via screenshots and downloads a CSV.
Why: 80% of enterprise software has no API. GUI agents unlock the legacy world text-based agents cannot touch.
4.) 3-Tier Memory & Eviction Engine
Build: A memory system with working scratchpads, episodic session compression and semantic vector search with automated TTL eviction.
Why: Infinite context causes OOM errors and hallucinations. Memory must be actively managed, scored and pruned.
5.) Multi-Agent Swarm with Typed A2A
Build: A 3-agent swarm (Researcher, Coder, Reviewer) that passes strictly typed Pydantic state objects, never raw text.
Why: Single agents fail on complex tasks. Swarms divide and conquer but raw text handoffs cause context drift.
6.) Sandboxed Code Executor
Build: A code-interpreter agent that writes and executes Python in an isolated E2B or gVisor microVM with zero internet egress.
Why: Letting an LLM run `os.system()` on your host machine is a catastrophic security risk.
7.) Inter-Agent Sanitization Proxy
Build: Middleware that intercepts messages between agents, stripping malicious instructions to prevent cross-agent prompt injection.
Why: Multi-agent systems create internal attack surfaces. Agent A can be tricked into hijacking Agent B's tools.
8.) Async HITL Approval Gate
Build: A workflow that drafts a $10k invoice, serializes its state and waits 3 days for a human to click "Approve" in Slack before resuming.
Why: High-stakes automation requires human accountability. Agents must know how to pause, wait, and resume cleanly.
9.) Self-Healing Coder Loop
Build: An agent that reads its own stack traces, parses the compiler error, rewrites the failing code, and retries up to 3 times.
Why: Agents that fail on the first try are useless. Agents that debug themselves are digital employees.
10.) Trajectory CI/CD Gate
Build: A GitHub Action that replays 100 agent trajectories and blocks the PR if the tool-call sequence deviates from the golden DAG.
Why: Grading the final text output hides the fact that the agent hallucinated an API call to get there.
11.) API-to-Browser Fallback Router
Build: A routing layer where if the primary REST API 404s or times out, the agent seamlessly spawns a headless browser to finish the job.
Why: The real world is messy. Agents must adapt to broken environments without crashing the workflow.
12.) Agentic Circuit Breaker
Build: A middleware tracking loop depth and tool-call recursion that instantly halts the agent if it enters an infinite spiral.
Why: A runaway agent loop is an outage with an invoice. It will burn $500 in API credits before you notice.
13.) Dynamic Context Budget Orchestrator
Build: A pre-flight assembler that strictly allocates tokens across system prompts, retrieved docs and tool schemas per request.
Why: Context overflow is the #1 silent killer of production agents.
14.) Shadow Trajectory Comparator
Build: A silent 5% traffic mirror that runs a new agent topology in the background, diffing tool-call paths and costs without affecting users.
Why: You cannot A/B test non-determinism safely in production without shadow mode.
15.) Public Architecture Teardown
Build: 3 published deep-dives detailing your swarm topology, state-machine design, inter-agent security and cost-per-task analysis.
Why: Top AI labs hire Agent Engineers based on their system design judgment not their ability to import LangChain.
Systems that prove you can orchestrate, secure & scale when autonomous agents touch your stack.
Most people watch tutorials. Builders ship systems.
Bookmark & Repost.
If you learn AI Agent Engineering, you will never be unemployed again.
I’d do this.
Stage 1: Agent Architecture & State Machines
- Learn: Directed Acyclic Graphs (DAGs), cyclic workflows, LangGraph, Temporal dot io, state serialization.
- Practice: Build a multi-step agent using a state machine that can loop, branch and pause without using a fragile `while True` loop.
- Why: Chat loops crash in production. Real-world workflows require deterministic state management and crash recovery.
Stage 2: Tool Use & MCP (Model Context Protocol)
- Learn: JSON-RPC, strict Pydantic schemas, MCP server/client architecture, dynamic tool discovery.
- Practice: Write a raw MCP server from scratch exposing a local Postgres DB, handling auth and scoped permissions without using a wrapper SDK.
- Why: Agents need standardized, secure hands to do real work. Bespoke API wrappers do not scale across an enterprise.
Stage 3: Computer Use & GUI Automation
- Learn: DOM parsing, Set-of-Mark prompting, coordinate mapping, screenshot diffing, Playwright/Puppeteer.
- Practice: Build a vision-based agent that logs into a legacy web portal, navigates menus via screenshots and downloads a CSV.
- Why: 80% of enterprise software has no API. GUI agents unlock the legacy world that text-based agents cannot touch.
Stage 4: Advanced Memory & Context Engineering
- Learn: 3-tier memory (working, episodic, semantic), vector eviction policies, context compression, token budgeting.
- Practice: Build a memory engine that automatically compresses a 50-turn conversation into an episodic summary to prevent context overflow.
- Why: Infinite context causes hallucinations and OOM errors. Memory must be actively managed, scored, and evicted.
Stage 5: Multi-Agent Swarms & A2A Handoffs
- Learn: Agent-to-Agent (A2A) protocols, typed state passing, supervisor vs. peer-to-peer topologies, consensus voting.
- Practice: Build a 3-agent swarm (Researcher, Coder, Reviewer) that passes strictly typed JSON state, never raw text.
- Why: Single agents fail on complex tasks. Swarms divide and conquer, but raw text handoffs cause context drift and hallucinations.
Stage 6: Sandboxed Execution Environments
- Learn: E2B, WebContainers, gVisor, Docker-in-Docker, resource limits, network egress blocking.
- Practice: Build a code-interpreter agent that writes and executes Python in an isolated microVM with zero internet access.
- Why: Letting an LLM run `os.system()` on your host machine or production server is a catastrophic security risk.
Stage 7: Inter-Agent Security & Guardrails
- Learn: Indirect prompt injection, tool poisoning, sanitization proxies, least-privilege scoping.
- Practice: Build a middleware that strips malicious instructions from an email before the "Email Reader" agent passes it to the "Database" agent.
- Why: Multi-agent systems create internal attack surfaces. Agent A can be tricked into hijacking Agent B's tools.
Stage 8: Human-in-the-Loop (HITL) & Breakpoints
- Learn: Asynchronous approvals, webhook routing, Slack/Teams integrations, durable execution pauses.
- Practice: Build an agent that drafts a $10k invoice, serializes its state, and waits 3 days for a human to click "Approve" in Slack before resuming.
- Why: High-stakes automation requires human accountability. Agents must know how to pause, wait and resume cleanly.
Stage 9: Self-Correction & Reflection Loops
- Learn: Critic agents, traceback analysis, retry with context, compiler error parsing.
- Practice: Build a coding agent that reads its own stack traces, rewrites the failing code and retries up to 3 times before escalating to a human.
- Why: Agents that fail on the first try are useless. Agents that debug themselves are digital employees.
Stage 10: Trajectory Evals & CI Gates
- Learn: Step-level grading, golden DAGs, trajectory distance metrics, LLM-as-a-judge for tool calls.
- Practice: Build a CI pipeline that replays 100 agent trajectories and blocks the PR if the tool-call sequence deviates from the golden path.
- Why: Grading the final text output hides the fact that the agent hallucinated an API call or skipped a safety check to get there.
Stage 11: Graceful Degradation & Fallbacks
- Learn: Circuit breakers, API-to-browser fallbacks, model cascading, cost-kill switches.
- Practice: Build a routing layer where if the primary REST API 404s, the agent seamlessly spawns a headless browser to finish the job.
- Why: The real world is messy. Agents must adapt to broken environments and degraded APIs without crashing the workflow.
Stage 12: Public Agent Teardowns & Portfolios
- Learn: Architecture decision records (ADRs), cost-per-task analysis, latency profiling, public benchmarking.
- Practice: Publish a teardown of your multi-agent swarm, detailing the exact cost, failure rate, state-machine design and eval methodology.
- Why: Top AI labs hire Agent Engineers based on their system design judgment and production reality not their ability to import LangChain.
Bookmark and Repost!
😭
🇩🇪 Audi used to do crazy stuff for advertising too!
In 1986, peak of German car industry, Audi made this legendary commercial! An Audi Quattro driving up a ski jump slope to show off its traction and 4 wheel drive.
That was a key moment and Audi dominated racing for years.
This was a kind of “Tesla flying moment”.
Why did they stop innovating and dominating?
Cars from that time still drive like a charm by the way.
Tobias retweeted
holy shit, opencode got hacked and it's officially fu*ked
'a large scale distillation attack, 1.4% of console users hit'
and people got drained, one user says his openai and anthropic keys, used only in opencode, ran up over $400 overnight, another replied $1,700. who compensates that?
every one of them had external providers linked, and it's still ongoing
this is the warning for all of us, build your own harness and keep your keys on your own machine, don't park them in anyone's console anymore
We had a security breach last night related to a large scale distillation attack that impacted 1.4% of OpenCode Console users
We have notified them all this morning via email. If you did not get an email you were not impacted, it was a subset of users who had external providers linked
The incident is ongoing so we cannot share more details at this point but we are working with Anthropic, OpenAI, and Google to remediate
Tobias retweeted
Gemini 4 Argon is amazing for knowledge work and coding - and it’s super fun to build with! We asked it to create this wizard chess game and it's so cool to see Argon's physics understanding & 3D modelling capabilities, makes me nostalgic for Battlechess on the Amiga - fun times!
Argon is rolling out today to Google Ultra and Pro subscribers in the @GeminiApp and Search, and Ultra subscribers in @antigravity! More info & full availability in the blog: blog.google/innovation-and-a…
Tobias retweeted
The moment you've been waiting 4…
Our new frontier model, Gemini 4 Argon, is rolling out now in the Gemini app to bring your most complex ideas into reality.
See how Gemini 4 Argon and our latest updates across the Gemini app help you get the most out of every prompt. 🧵
Tobias retweeted
OpenCode 2.0
anoma.ly/notes/opencode-2-0/
Tobias retweeted
Most people don’t actually want to own luxury bullsh*t. They just want free time to do whatever the hell they want without being told what to do.
Nailed it
It’s been very hard for ai researchers and companies to realize most of the public are not entrepreneurial and do not think of creating or building things or connecting systems or anything beyond casual consumption. It’s a bummer but karpathy’s numbers are solid here. This is why social media and shopping focused consumable ai will do so well because it’s built by people who understand that most people just want to scroll and shop and consume, and laugh, not create, build, innovate. Of course, we want to empower more people to do those things, but the Internet empowered that and AI has already empowered that the last two years alone. Hopefully more people will utilize these tools and incredible opportunities to lead with Superintelligence. most won’t so the few who do will win and widen the gap.
Tobias retweeted
This is a good way to understand how early we actually are in AI diffusion right now. The vast majority of people are using AI as a chatbot for simple interactions that approximate next-generation search. This is super useful but dramatically underrepresents the full capability of AI.
On the complete other end of the spectrum, agent swarms represent possibly the ultimate form factor of what agents will do inside of our workflows. Math, cyber, bio, materials science, software engineering, finance, and any other field that has valuable tasks that can be executed in parallel where compute can be applied to accelerate work all have this characteristic.
Less than a fraction of a fraction of the population uses AI in this way right now, due to a mix of the capabilities of models relative to the user's domain, the cost of doing these types of workflows today, the harnesses available to do this style of AI work, and the availability (or lack thereof) of frontier models to the general public.
There will also be a huge amount of change management necessary to get most areas of work -beyond software and AI research- to benefit from this style of agents. Companies need to upgrade their workflows, data needs to integrated and migrated into modern platforms, and companies will need to work through compliance, governance, and security frameworks to pull it off.
But this all represents opportunity right now. A lot of this is where the applied layer of AI will really shine in the enterprise because of the domain specificity of this type of work. You can get an early sense of where things are going and the opportunity ahead by seeing through the funnel that Andrej lays out.
Dusting off this tweet from April because this gap in shared understanding of LLM capability is *widening*. It's now less "two groups of people speaking past each other" and more a sharp funnel.
- Napkin math somewhere around 6B people (~75% of the population) have barely come in contact with LLMs at all.
- Around 1-2B (~20%) are casual and infrequent users of free-tier ChatGPT-like products. This group sees derpy chatbots and treats them a bit like a better Google search, a writing aid, or etc. Maybe an agent tries to book you a flight. I have non-tech friends who (reasonably, imo) say they have not much use for it at all. Even many professionals outside of math&code are in this tier. For example, execs and many other functions spend a lot of their time talking to other people, so while they understand what is happening intellectually, it is still second-hand and a bit abstract.
- Now we get to professional use of frontier-grade LLMs in math&code. Somewhere around 20M people (0.2%) see first-hand that large, complex projects that used to take them weeks/months can now be completed by agents with a prompt. Building apps, copying apps, translating apps, decompiling apps from binaries... This has all happened very quickly and recently - less than 1 year ago, I was writing code manually by hand, typing memorized computer code commands into a code editor character by character, occasionally pressing Tab to autocomplete a little chunk of code.
- And finally we get to the ~5,000 people (~0.00006%) with access to frontier-grade systems internally. The external world has seen the preview. It looks like swarms of thousands of agents collaborating over weeks on software mega projects: minting zero days, running cyber attacks and defenses at machine speeds, discovering new science, advancing the frontier of mathematics. Things that would have taken top professionals in the industry years of work. Meanwhile, human review and comprehension are starting to fall behind. For example, people are still involved in the "archeology" of the OpenAI-HF incident from many months ago. Mathematicians may be poring over the 722 manuscripts on frontier mathematics for a while.
The funnel is driven by a combination of factors:
1. The impact scales with ambition, problem size, and horizon. A question with a paragraph answer barely stresses the system. You need a reservoir of big, difficult problems that you really care about. This aspect drives the consumer / professional dimension of the funnel.
2. The jaggedness of the system (which I have written about a lot separately). Capability peaks in domains that are digital, verifiable and economically valuable. This is because LLM capability emerges from reinforcement learning on verifiable rewards on a curated environment mixture driven by revenue potential. This aspect primarily drives the area (e.g. math&code) dimension of the funnel.
3. Access. Free-tier, paid-tier, internal.
So this is the weirdness of the moment. The general public has mostly not interacted with these systems. When they have, it looks like a derpy chatbot. The majority of professionals still see only a modest uplift. And a small sliver of professionals are experiencing the vertigo of the curve going vertical. And it is all happening at the same time.
Tobias retweeted
Agents spawning other agents, agents operating in the background on every task, and agent swarms will consume 1,000X more tokens than people prompting agents 1 at a time.
The usage of AI as a chat system that can only work for you at the pace that you can prompt it will seem like a relic within a year or two. The vast, vast majority of tokens will be consumed by agents that are just doing continuous work for us in the background and in our workflows.
This is why we’re still so early in the curve of AI agent adoption right now, and how much more compute and infrastructure build out will be needed to sustain this.
Tobias retweeted
Grok @Bot will use whatever achieves the best outcome for users.
Simple questions will route to small, fast models. Questions with complex answers will route to large models.
Tobias retweeted
Not my usual style to post half naked photos, but what the hell... 🤣
Also 49, no AI, no photoshop, pure human labor.
I encourage people to compete on health, not wealth rankings or lambos.
More health, more family = true wealth, and happiness.
Not bad for a 49-year-old! One year away from the half-century mark.
Great birthday workout with my brother and trainer, Ossie.
Feel free to DM me for tips on training, health, and wellness. And catch me October 11 at the More Life Summit: morelifemvmt.com