PhD mathematics; scientific software architect; mountain biker / runner. Opinions are my own and do not reflect any associated entity.
Colorado
Joined July 2022
- Tweets4.3K
- Following510
- Followers1.4K
- Likes5.5K
Colin Roberts retweeted
We cut Multifactor down to one thing: sharing an online account by sending a link.
Pick what each link can do, watch every action live, revoke in one click. Your password never leaves.
Built in our second Lock-In Week. Try it (Hacker News for now): lite.multifactor.com
multifactor.com/blog/introdu…
Colin Roberts retweeted
😅😂
9th edition of taco bell 50k ultra marathon is happening this Saturday October 3rd in Denver Colorado.
This is one of the weirdest marathons you will see. Runners have to stop at all 10 taco bells. They must eat a menu item from 9 of 10 taco bells. By 4th stop everyone must eat a chulapa supreme or crunch wrap supreme. By 8th stop everyone must eat a burrito supreme or nacho bell grande. And they have to finish it under 11 hours. Race starts at 7:30 am
#food #Marathon #colorado – at Denver, CO
Colin Roberts retweeted
yep this is the way.
i built fitness-tracker.me for exactly this. photo upload, LLM, but also includes historical tracking, trends and charts, protein/fiber/caloric deficit targets.
its totally free, built it for my own usage, but you might like it
Colin Roberts retweeted
You gave your agent access to your X account and asked it to summarize your DMs.
How do you know it read all of them? How do you know it didn't also reply to posts, or make its own?
Password managers handle authentication, not authorization. OAuth and MCP need the website to play along, and your regional bank isn't shipping an MCP server.
We built Checkpoint: scoped access to any account, enforced at the network layer inside a TEE, delivered as a hyperlink anything can open.
multifactor.com/blog/checkpo…
Colin Roberts retweeted
Performance Engineering at TigerBeetle, @maxikuschewski 's guest lecture at @TU_Muenchen:
01:59 Understand your Workload
05:04 Performance
11:36 Tail Latency
21:52 Algorithm Engineering
28:02 Async I/O in TigerBeetle
35:21 Object Storage Clients
youtube.com/watch?v=-p-Xm7VE…
Colin Roberts retweeted
Object storage SDKs are often careless with blocking syscalls; game over for single thread. They introduce dependencies, spawn their own threads, and allocate/deallocate memory. Not TigerStyle.
So, we implemented TigerBeetle's own object storage.
Colin Roberts retweeted
For an AI agent to do real work (pay an invoice, pull a statement, check an order), it has to get through a login page first. Login pages are built for humans: multi-step forms, one-time codes, passkeys, captchas, layouts that change overnight.
Neither naive approach held up for us. Scripts break on every redesign. And letting an AI improvise with real credentials on a live account is not something we were willing to ship.
At Multifactor, we split the problem:
- An AI recon agent explores each website ahead of time, with no customer credentials involved, and produces a deterministic, auditable login flow.
- Every flow is tested in our lab with our own agent credentials, then manually reviewed and released individually. When a site changes overnight, the flow self-heals in the lab and re-enters that same review gate.
- At login, Checkpoint replays that flow inside a trusted execution environment. No AI in the loop when your credentials are in play.
- Success is verified with evidence: signals the login itself caused, instead of waiting and hoping.
Full write-up by our founding engineer Pierre Ortega: multifactor.com/blog/teachin…
Colin Roberts retweeted
6 months ago, we met a team that was remarkably well aligned with our agentic security vision.
Today, we're joining forces.
We've officially acquired Keypo, Inc., and its founder Dave Blumenfeld is joining us as Founding Engineer, bringing his work on secure enclaves directly into our platform.
Colin Roberts retweeted
We’ve had on the order of 3 memory bugs in 6 years of TigerBeetle. None RCEs.
On the other hand, our own simulators have proactively found hundreds of (devastatingly catastrophic) distributed systems correctness bugs per year.
Given how hard TigerBeetle’s domain is, in terms of mission critical financial transaction processing, I’ve never for one minute believed that writing TB in a memory safe language such as say TypeScript would somehow magically (!) make any material impact compared to the 100x correctness multiplier of TigerStyle.
That’s because—rather than fall for the fallacy of composition, i.e. to see distributed correctness as a language problem—TigerStyle instead takes ultimate responsibility for the “end to end” correctness of the distributed system as a whole.
Per systems engineering, correctness is always a systems design problem. For example, how to build a reliable whole, (especially) out of unreliable parts, such as broken firmware, bitrot, programmer error etc. In other words, application of the end to end principle.
But when you TigerStyle the design in this way, the world of systems engineering also completely opens up to you and changes how you evaluate systems languages (now things like “power to grammar ratio”, or explicitness, checked arithmetic and precision become more critical and valuable to you).
Of course, it is harder to care about correctness, to take responsibility for correctness end to end.
Yes, you’re forced to begin to worry about the more serious concerns, starting with the basics of static allocation, explicit limits, assertions, deterministic simulation testing and moving to more advanced topics like protocol-awareness and storage fault-tolerance.
But then again, TigerStyle is such a force multiplier, that you achieve mission critical quality, and in less time and with greater velocity.
If you’re tired of production issues, and if you want to “engineer your engineering”, I would encourage you to lift up your thinking to the level of systems design and end to end correctness.
Start thinking about your methodology and begin embracing TigerStyle.
tigerstyle.dev
Colin Roberts retweeted
seriously, working with AI is MISERABLE for one and only one reason: having to re-explain the same thing
"oh yeah this new session obviously doesn't know what proper case trees are, so let me explain it for the 5000th time in my life"
I'm tired
AGENTS.md doesn't solve this because it is impossible to fit the entire domain knowledge without nuking the context - it would be 1m+ tokens worth
RAGs don't solve this, the agent won't search unknown unknowns
SKILLs don't solve this unless I keep like a collection of 1750 skills with specific cuts of domain knowledge for each possible subset of my domain that I might need in a given chat, but that's a lot of manual work
recursive LLMs or whatever don't solve this for the same reason, you can't dump a domain book and expect the AGENT will magically guess that it is supposed to search for a specific bit knowledge. unknown unknowns
fine tuning doesn't solve this (OSS models suck and OpenAI / Anthropic gave up on user fine tuning)
I honestly think a good product around fine tuning on your domain would be a major hit and an underdog lab should take this opportunity
Colin Roberts retweeted
⚡️ Zig 0.16 is out.
And the new I/O model is a huge shift.
• Swap implementations (threaded, evented, etc.)
• Write code that looks blocking but runs async
• Composable like allocators
ziglang.org/download/0.16.0/…
#zig #ziglang
Colin Roberts retweeted
There's something about @olgazaghen's review of Jacobi Fields that feels like Denoised Image Fields from Pixel MeanFlows @ZhengyangGeng @Lyy_iiis
Colin Roberts retweeted
Anyone interested in:
- Rust / Zig / C++
- Mathematics
- Machine Learning
- Scientific Computing
Come join us here! discord.com/invite/qEwPr3GMP…
Colin Roberts retweeted
TigerStyle.dev
A new, visual home for TigerStyle -- the software engineering methodology developed by TigerBeetle to produce safer, faster software in less time.
"Do the hard thing today to make tomorrow easy."
"If you want to be remembered, be remarkable."
Colin Roberts retweeted
Just listened to Peter Thiel complain about how hard it is to have mega companies in Europe and how everyone wants to sell and go on vacation. I am glad Europe exists as a counterpoint to the rat race that is the US. I absolutely love how life can be slow here, how money isn’t everything. How in much of Europe no one cares if you are a venture capitalist with a gazillion dollars to spend. How people don’t need to work themselves to death to prove a point. How on a Wednesday afternoon you can find people just enjoying the sun and not fretting about deliverables. How trains work. How guns are not everywhere. How road rage doesn’t not immediately lead to someone shooting you in the chest. How cities are walkable. How people are not obsessed with marriage and who pays on dates. How little purity culture there is. How people are not terrified of a naked body in public. Because in the end after you make all that money, after you build that mega company, what is it all for? What is all your success worth if in this brief life you cannot enjoy the things that make life worth living, the little things, the lakes and rivers and oceans, the mountains, the colors of nature, the taste of food, the smiles of humans who are not all judging each other based on which patch of land they own, which piece of equipment they have, and what imaginary figures they have in banks? What is human progress for if not for the enjoyment of this one brief life? And yes Europe has its problems. A lot of them. But it would be a disaster if Europe became like the US.
Colin Roberts retweeted
ARC Sandbox is a new effort to iteratively automate experiments for discovering solutions to @arcprize.
I've not had the time to explore all the ideas myself, but for the first time it feels like I have a team of researchers working together thanks to Claude/Codex/OpenCode using MLX (in the spirit of solution efficiency).
At the core is a collection of papers (docs/references.md) that inspire ideas, then we focus on generating new experiments by (1) exploring concepts, references, and past experimental plans/implementations (2) create a detailed plan for the experiment with derivations, pseudocode, limitations, etc. (3) implementing the new experiment (4) evaluating the generated solutions (5) incorporating feedback back into the plan and refining the implementation.
My responsibility becomes ideating (using gpt-5.2) and updating papers to then advise new agents on creating experiments. An initial push in the experiment direction can also help avoid repetition and encourage orthogonal directions.
Results are promising! I expect the the workflow and implementations will also improve with time, and I plan to share progress updates as this work evolves.
