Dima

Steak holder, stuff engineer. Building @xmemory_ai. Host of the @SysDesignMeetup. CS generalist, ex. competitive programmer.

[xmemory @xmemory_ai]

Agents should get better with experience. We teamed up with @tolokaAI to measure that.

Toloka evaluated xmemory on closed datasets. No peeking at the data. No tuning to the benchmark. Just learning from failures and measuring what improves.

📈 The results: 26% lower token costs when xmemory replaced document search as the policy store. And adding lessons from past runs lifted scores by 4.8 percentage points on held-out tasks.

💡 It means with memory, a mid-size open-weight model scored 0.75 – beating Sonnet 5 at 0.68 and GPT-5.6 Terra at 0.69 without memory.

This experiment is just the beginning of what memory can unlock. We see even bigger gains on longer-horizon tasks, where every lesson has more chances to pay off.

Read our blog post below!

[Dima @UniqueDima]

Yay for @yonasbe!

[Yonas Beshawred @yonasbe]

35% speed up with 5 minutes of work!

If you’re still paying for the default GitHub Actions runners, there’s no reason not to try @starslingdev 💫

[Dima @UniqueDima]

Ref. https://t.co/kXRf6dFHul and Josh himself!

[Daniel Svonava @svonava]

So, squares huh? @ojoshe @UniqueDima @naveen23k

[Dima @UniqueDima]

Serious question out of the blue: Do we even need Linux for servers these days?

A friend sent me https://t.co/MMDqEu9KbU over the weekend. I looked into it and could not form an opinion on whether it will or will not fly. From first principles, it should. The reality check, however, is that it's a few years old already, and not taking the world by a storm.

Then I thought deeper. Golang, with all my disrespect to it as a language, is increasingly popular. And it's goroutines-first, i.e., cooperative multitreading is its building block.

Most Rust production-grade code uses tokio these days. Or something like it. So, Rust code is also cooperatively multithreaded by design.

Also, look at Nvidia GPUs. We, the world, have proven that a small set of chips can run the planet. No need in a huge variety. Apple's M1-M5 chips are another proof.

Besides, Google was experimenting with turning old Pixel phones into the CPUs that power a datacenter.

So, compute-wise, do we even need linux? Why can't we take one well-designed system-on-chip ARM, and give it a tiny wrapper on which a Golang runtime or a Rust tokio runtime can flourish?

And it can be a thin real-time OS, which Linux is definitively not.

Seriously, seen in this light, most of POSIX can, and arguably should, be treated as legacy. It's a source of unwanted synchronicity, and it's a source of bugs, after all. A tiny kernel, a blob storage, and a well-optimized async-friendly massively parallel network controller is all we need.

So, from first principles, I can see a competitor to AWS emerging, where at a fraction of cost we can run Golang code or Rust code with increased reliability. The code may have to be implemented to use special libraries, and built with a dedicated toolchain, but this does not appear to be a problem in the day and age of agents.

What am I missing? As in, aside from the legacy "too big to fail" argument, why should we not expect something like this to happen in the next year or two?

[Dima @UniqueDima]

As we discovered today, "please call the manager" when the staff is acting silly is literally the same as turning model reasoning on and retrying the prompt.

The very same prompt in most cases, I should add. Just against a reasoning engine with a larger context window.

[Dima @UniqueDima]

"People are irrationally hesitant to show their prompts", "do you go back and correct typos in your prompts?" (c) Peter Steinberger

Dude, sometimes I prompt without spaces to keep the room quiet.

Pasting an image and saying `dafaq` is a marker by which OpenAI and Anthropic can probably trace my presence across platforms.

And if my keyboard layout is сщташпгкув цкщтпдн and I notice it mid-paragraph, I would generally just continue ещ ензу фы шы, because why waste precious human time on something agents will handle in a split second with zero effort?

People who hesitate to share their prompts have it backwards. "It's a boomer thing", as Peter rightly says.

Thx for the event yesterday!

[Dima @UniqueDima]

Someone really needs to do to git what uv did to python.

I'd chime in, but am oversubscribed now :-(

[Dima @UniqueDima]

@Nil053 But you still read stuff online, right? :)

[Dima @UniqueDima]

This deserves an entirely new browser, and I, for one, would try it out if it's open source.

[Nil @Nil053]

Idea: Vibe Browsing

Ask your agent to make a browser extension that does nothing.

Install it, pin it.

It's your way into vibe-coding how you browse.

Any time you don't like something, ask your agent to change the extension code and click "reload".

- Hate Shorts? Gone
- A site blinds you? Dark mode
- Missing a feature? Added

[Dima @UniqueDima]

Oh well. Fix it, dear agent.

[Dima @UniqueDima]

[1] https://t.co/AnpBCqcd4P

[Dima @UniqueDima]

A gentle reminder that we're hosting the Building with Context event on #SFTechWeek, in the evening of Thursday, October the 8th.

It's already packed with good engineers, yet Staff++ folks are very much welcome. Link in [1].

[Dima @UniqueDima]

2021: Why can't we solve machine translation?

2026: Wispr Flow is so good at trying to make my speech make sense as text that it literally parsed my three lines long Russian instructions as Ukrainian.

I don't speak or understand Ukrainian. I've translated that resulting text back into Russian. That's exactly what I wanted my agent to do.

"Accidental translation" in 2026 is literally better than bleeding-edge translation in 2021.

Cross-language understaning is a commodity now. It is cheaper to translate in 2026 than it was to understand штсщккуседн-сщташпгкув-keyboard-layout text in 2021.

[Dima @UniqueDima]

Inspired by a recent conversation.

I can somewhat understand copyright on authored texts when it comes to pieces of art, also known as literature. But I can't understand no-AI-training clauses or README demands that models not read the code.

So you are publishing a piece of code that you are not allowing models to train on? And I immediately think of consumer tools, such as CLI / TUI utilities.

Well, if it's useless code, it's useless regardless.

And if it's useful, a *lot* of people are using this code.

It's easy to use programmatically.

It's easy to use when it's run on a remote machine.

It's easy to use when it's 100% sandboxed.

It's easy to put under `tmux` and `asciinema`, which I do routinely.

And presumably your tool has decent help. And forums. And tons of data on how best to use this tool, and how it is used.

And then you say "don't read this source code with AI"? Seriously?

It would take single-digit dollars — LITERALLY SINGLE-DIGIT DOLLARS — to blackbox-sandbox your tool, DOWN TO ALL OS CALLS IT MAKES, and to ask a few AIs to read its help, to read the forums, and to come up with a fully functional demo of the features people actually use.

And then another, I dare say, SINGLE-DIGIT HUNDREDS OF DOLLARS to reimplement those features from scratch.

Without a single glance at your tool's source code.

So, if it's a useful tool that people need, you're just sitting on a ticking time bomb if you choose to sit on its otherwise-open code as a dragon hoards its treasure. The rebuild happens either way; sitting on the source just leaves you outside the process. In fact, everything I'm saying above applies to closed-source software too.

With proprietary, closed-source software, by the way, the difference is enforcement. The same black-box rebuild applies, but lawsuits are only plausible where there's real money on the table; real money that could be won. And a large corporation behind those lawsuits, with an army of lawyers and with tons of money at stake.

But a low-key developer, or a community of low-key developers, who have been maintaining some tool(s) over the years? Sure, you could argue you've invested a few hundred hours into your baby tool; perhaps a few thousand hours. You're not going to court for it. Moreover, you're worth more as the expert who steers the rebuild than as a pile of source code you won't let models read. You can LITERALLY approach big AI companies and offer that expertise.

After all, a great-to-use tool with wide adoption — wasn't this what you were working for all this time, as enthusiasts? During "peace times", i.e. assuming there's no "AI war" for training data, who cares about the code, except perhaps security researchers? You built something many people use, great. Now the Universe is offering you a chance to make your own impact larger, by a wide margin. All you need is to say yes and collaborate.

And yet quite a few people choose to turn their egos on instead, and "protect their code" from being "used by the AI to improve". As if they would have protested if some professors had showcased some important concepts to students based on their code or something.

Don't get me wrong, there are quite a few prominent critics of Big Tech. I'm sure I can steelman their arguments too. RMS, for instance, is largely right, IMHO, that the death of truly open software is what created this siloed monopoly-run Big Tech, as opposed to Hypertext FIDO 2.0. But copyleft was about keeping code free to use and modify, not locking it away from machines that read it. And the cypherpunks community, while it has brought many useful tools and privacy wins such as PGP, is largely gone today. So yes, it's not difficult to steelman that criticism.

But why not join forces? Why not view the AI as the force of nature, as the greater good, that we should all collectively leverage for a better civilization? Why instead choose to resurrect the luddites movement and fight against?

[Dima @UniqueDima]

So we have a solid addendum to open-weights models thanks to Mistral, or is this a wrong way to read the news?

[Dima @UniqueDima]

I wonder that is saying "for Sam's sake", "for Dario's sake", and "for Elon's sake" in prompts would yield better results.

[Dima @UniqueDima]

Yes.

[Nil @Nil053]

For someone who derives meaning mostly from creating, it's weird that I spend essentially no time thinking about what to create.

A good way to model this is to imagine that there's an external entity, call it Intrinsic Compass, deciding what's worth making and communicating it via implanted thoughts. I'm just an IC proxy - I don't quesiton it, I just work hard to make it happy.

This model actually extends to other aspects of my life. For example, I mostly communicate with my friends by sharing my (the IC's) creations, be it blog posts, songs, side projects, etc. And then they complain that they want to know more about ME instead. And I think, "me? I'm just a proxy, why would you be interested in the proxy?"

I reflect on this as I consider becoming a founder. A good founder thinks carefully about what to create, coldly optimizing for expected value. The IC would need to be jailed.

[Dima @UniqueDima]

I’m far from a believer that the US actually has freedom of speech.

But on a relative scale it most definitely does.

Just imagine it being problematic for an ordinary citizen to look up health condition details, and notes on surrounding events, of their leaders.

After living in the US for 15+ years I literally can’t. Knowledge is power, and say what you want about propaganda and doublethink, access to verifiable facts and the desire to seek them is still enormously valuable when beyond the critical mass of the population has it.

And yet hundreds of millions of people worldwide, today, are routinely avoiding even thinking of some possibilities, correctly assuming it’s a potentially punishable “thoughtcrime”. Worse part is, those people have a credible reason to believe it actually is punishable.

[Dima @UniqueDima]

Thoughts after sleeping on a recent incident where a major bank handed its customers' passports, selfies, and transaction histories to scammers who asked from a government email address. I can now phrase my position even stronger.

TL;DR: The problem is not just what the regulators and the governments are doing, although they are doing a lot. The problem is that they are too powerful, and no one can stand up to them and say "enough!"

This was not the case ten years ago. It was, five years ago already. I've seen it in DeFi / Web3, and it's just too obvious with ordinary banking.

Banks want to do business. It's too good of a business to be a bank. And to be a bank you have to be friends with the regulator. Regulators, in turn, want more and more bank friends, as long as they play by the rules. And the banks do. Of course they do.

What we lack is a force on the side of We, The People strong enough that some bank cracks under the pressure and openly states:

• Screw you, governments and regulators.
• Your rules are NOT protecting the customer, they are making it WORSE.
• You want more control, but it's just PRIVACY INVASIONS, and they WILL backfire.
• So yes, we, bank X, ARE WILLING to serve customers in your jurisdiction.
• But we will open OFFSHORE accounts for them, and ship cards by international mail. Perhaps under different names too, which is perfectly legal: the name on your card need not match your passport.
• Not to evade your rules. To PROTECT our customers.
• And now, dear regulator, PLEASE WORK WITH US. We can help you draft good regulation and establish guardrails, so that no regulator, YOU TODAY OR THE NEXT ONE TOMORROW, could stop putting bank clients' interests first.

There is literally NO SUCH BANK. And this is the problem.

I'm sure tons of founders in banking would LOVE to say this. They are already billionaires, so if such a move had a real chance, I would expect someone to have tried it by now. But the governments and regulators are too damn powerful. Power corrupts, and absolute power corrupts absolutely. When it comes to financial overseeing, our overlords are corrupted.

And I am saying this as an ordinary, law-abiding citizen. I do support KYC and AML. I pass them regularly. I want more people to pass them when buying a house next to mine, a yacht, or a large stake in a startup. Hell yeah, track the source of income.

But the system should be designed bottom up. Not "we, the Regulator, must see everything", but "here is the list of acceptable documentation: provide it and you are GUARANTEED to pass KYC/AML". Tax forms and income statements, for instance. Not one's home address, for God's sake, not one's net worth, and definitely not one's family and family businesses.

A trivial change, really. And finally: make it unconditionally legal to bank in a different jurisdiction. Today, banks in plenty of countries refuse citizens of certain other countries, only because those countries make it too damn hard. It should be perfectly fine, as long as one pays their taxes at home. Fierce competition would follow, and transfers would take seconds for EVERYBODY, because commerce works. Top-down suffocation does not.

The Web3 folks have already built the rails. Seconds-long settlements for fractions of a penny, at scale, with receipts and proofs attachable to any transaction, so that the parties CAN CHOOSE to show them, decrypt them, or ZK-prove something about them without revealing the rest. With an on-chain USD[C] account and Web3-friendly payroll and brokers, one's tax forms are a) fully automatic, b) perfectly auditable, and c) impossible to make a mistake in. Yes, impossible. Some operations, designed right, leave no room for a mistake, and getting paid while the tax authority knows about it is one of them.

We had a chance to live in this world by 2024. Or so it seemed in 2016. We blew it up.

But if enough sane powerful people say "enough!", the tide may well turn. That is what I would love to see.