ClintEhrlich

You could definitely build a coding agent that way, and it sounds like you've done it. We store the conversation history because:

1. In our use of coding agents, we find that there are often things referenced earlier in the conversation (API keys, endpoint addresses, feedback to the agent, etc.) that it's useful to have persist.

2. This is a general-purpose LLM memory system, which we've just used here to build a coding agent. But it is also designed for personal assistants, legal LLMs, etc.

By construction, individual summaries are not typically large enough to overload the context window when expanded.

The reason that the volume is potentially arbitrarily large is that one sub-agent can call lcm_expand multiple times - either vertically or horizontally. But that's a process that occurs gradually as the tool is used repeatedly.

This has not been a problem in our testing, but if it were a problem it would be easy to prevent sub-agents from invoking lcm_expand once their context buffer has reached a specified threshold.

Thanks for the kind words! Looks cool!

Hi NWU,

We don't have any other materials yet, but let's see if this lands for you. I can run you through a couple simpler versions of the system, why they don't work, and how that informs our ultimate design.

The most basic part of the system is "two layers". Layer 1 is the "ground truth" of the conversation - the whole text the user sees. Layer 2 is what the model sees, i.e., the active context window.

In a perfect world, those would be the same thing. But, as you know, context lengths aren't long enough for that, so we can't fit everything from Layer 1 into Layer 2.

So instead we keep a "pointer" to the appropriate part of Layer 1 in Layer 2. That pointer takes the form of a summary. But it's not a summary designed to contain all information. It's more like a "label" that makes sure the model knows where to look.

The naive version of the system would allow the main model to expand Layer 2 summaries by importing all of the underlying data from Layer 1. But this doesn't work well, because then you just end up re-filling the Layer 2 context window.

So instead you let the main model clone itself, the clone expands the summary in its context (and can do this for multiple summaries, transforming each into the original uncompressed text), and then the clone returns whatever information the main thread requires.

Where this system would not fully match the capabilities of RLMs is that, by writing a script that calls itself e.g. thousands of times, an RLM has the ability to make many more recursive tool calls than can fit in a context window. So we fix that using operator-level recursion, i.e., we give the LLM a tool, map, that executes arbitrary recursion, without the LLM having to write a custom script to accomplish that.

Hope this helps!

- Clint

Our system uses sub-agents as a core part of its architecture.

That terminology can be confusing, because in other cases (and sometimes in our own architecture, like when executing thousands of operations via MAP) a sub-agent may be a smaller model given less complex individual tasks.

But the core mechanism we use for simulating unlimited context is to allow the main model to spin up instances of itself (sub-agents) with the previously summarized portion of the context expanded into its full, uncompressed state.

Expanding summaries into full text in sub-agents rather than the main thread is a critical part of our architecture, because it prevents the main context window from filling up.

Just passed this on to my co-author who is working on the plug-in. Really appreciate the suggestions!

We will probably ship a fairly basic version to start, but I think there are a lot of cool things that can be added.

Oh and to be clear YES you can try it!!!

Just bring an API key. :)

github.com/voltropy/volt

Thanks for the kind words.

Yes, we think there is a ton of low-hanging fruit from taking lessons from OS/PL theory and applying them to LLM tooling.

This is our first contribution in that direction. There will be more!

Yes, that is actually the next thing we are shipping!

We have heard from a ton of OpenClaw users that the biggest barrier to them getting everything they want out of their agents is that memory is not a solved problem.

LCM could be a great solution to that. Stay tuned -- will ship it ASAP.

Hi, I'm Clint, one of the co-authors of this paper.

I'd like to quickly summarize what is different about our approach and why it matters.

Our work was inspired by brilliant research done at MIT CSAIL on "Recursive Language Models" (RLMs). One of the controversies has been whether these models are just a formalization of what agents like Claude Code already do vs. whether they bring new capabilities to the table.

By outperforming Claude on the major long-context benchmark, we provide a strong signal that something fundamentally new is happening. (In other words, it's not "just Claude Code" because it demonstrably outperforms Claude Code in the long-context regime.)

Where our contribution, LCM, differs from RLMs is how we handle recursion. RLMs use "symbolic recursion" -- i.e., they have an LLM write a script to recursively call itself in order to manipulate the context, which is stored in a REPL. This provides maximum flexibility... but it often goes wrong, since the LLM may write imperfect scripts.

LCM attempts to decompose the recursion from RLMs into deterministic primitives so that the control flow can be managed by an engine rather than left to the whims of the LLM. In practice, this means we replace bespoke scripts with two mechanisms: (1) A DAG-based context management system that works like paged virtual memory, except for managing conversations and files; and (2) Operator-level recursion, like "Map" for LLMs, which lets one tool call process thousands of tasks.

An analogy we draw in the paper is the evolution from GO-TO statements (of Dijkstra's "Considered Harmful" fame) to structured programming. RLMs are maximally expressive, but all of that power comes with the risk of things going awry. We have built a more mechanistic system, which can provide stronger guarantees when deployed in production with today's models.

Happy to answer any questions! Thanks for taking a look at the paper!

Thanks for the questions. I'll make sure to expand the FAQ when I get a chance.

1. The USD that you can use to unlock USDf (forked dollars/digital gold) is limited to deposits at commercial banks and credit unions. The public does not have access to electronic base money, so it's not included.

2. The quantity of USDf you can unlock is based on your historic bank balances, as verified during a defined window in the past. New credit money created after that point in time doesn't affect those past balances. To the extent that the owners of new USD wish to acquire weight for them, they'd need to purchase USDf, increasing its value. That is sort of the whole point: if governments keep inflating their money, the "forked" version with guaranteed scarcity will gradually increase in value.

3. USDf will trade at a different value than USD. At first, a much lower value. The idea is for them to be employed in a hybrid unit of account, USDw, where 1 USDw = 1 USD + 1 USDf.

A crypto-weighted dollar (i.e., USDw) will trade at a premium over a USD, since it is a USD + cryptographic weight. Think of it like a stablecoin that also comes with Bitcoin-like inflation protection. However, it's also possible to use USDf as an independent asset, and that will be convenient in use cases where transferring USD on existing payment rails isn't practical.

There's two nuances I would respectfully suggest that you're overlooking.

First, KRNC is designed to be employed as a supplement to USD. Most transactions would be executed with both USD and a corresponding blockchain asset. Technically, this is a digital analogue of the "symetallic standard", in which base money is comprised of both gold and silver in a specified ratio. The point is risk diversification: if fiat money implodes, or if crypto fails, you aren't wiped out.

Second, the concept of "intrinsic value" is misleading/confused when it comes to money. Things that trade at their consumption/production value are not monetized. Treating something as money involves attaching symbolic value to it: accepting it as proof of goods or services rendered in the past, and as a token that can be used to acquire goods or services in the future. Even gold would lose most of its value if it were suddenly priced based only on demand for use in industrial applications.

Money has always been valuable because everyone else treats it as money, whatever it is. It's a Schelling point that enables abstracted barter. Nothing less, nothing more.

That is precisely the problem we're trying to solve: allowing the public to take advantage of blockchain-based digital scarcity without anyone being forced to invest in speculative assets.

I'm actually trying to convince people that, if we accept the premise that blockchain technology can create digital gold (as the market has, to the tune of hundreds of billions of dollars) then we should harness that digital gold to protect the value of the money that everyone already owns, rather than setting the world on fire by launching new currencies that function like pyramid schemes.

I have no dog in the fight, and wish "enormity" had never developed a normative undertone, but I strongly disagree that said usage is anything close to archaic.

If you Google "enormity," the dictionary definitions it displays before the results are: 1.the great or extreme scale, seriousness, or extent of something perceived as bad or morally wrong. "a thorough search disclosed the full enormity of the crime" 2. a grave crime or sin. "the enormities of the regime"

Merriam-Webster claims this is not the exclusive usage, and that enormity can mean "immensity" without normative implications when the size is unexpected. But the very example it cites, from Steinbeck, involves the "enormity" of a situation in which a fire was started.

That said, I agree that "enormousness" is an awkward word, which I do not use. I'm left to ponder the enormity of my own pedantry.

Not quite.

Enormous : enormousness :: huge : hugeness.

Enormity = Immense scale of evil (e.g., the "enormity of the holocaust")

On second reading, I have but one trivial gripe:

"Enormity" implies moral reprobation, so it's a poor way of describing the significance of a computational discovery.

I've been waiting for Scott Aaronson to put all of this into perspective since the first leaks about Google's quantum supremacy started appearing in popular media.

He has exceeded my expectations with this post, which cuts through all the hype to communicate exactly what the results of this experiment mean for the field. It's worth reading and sharing.

Lead author here. This paper does a few things that HN may consider interesting.

1. It formalizes new security vulnerabilities in Bitcoin and other cryptocurrencies. Notably, it shows that all of these protocols repeat a simple statistical error that was introduced to the literature in the early 2000s.

2. It demonstrates how "honest majority" axioms can be replaced with a more rigorous formal method, which incorporates techniques from game theory and microeconomics to prove security from first principles.

3. It applies biological models to trust-minimized networking. By replacing handicap-authenticated signaling with cue-authenticated signaling, it obtains an exponential improvement in security and performance.

4. It proposes a cryptographic twist on the gold standard, which can deliver all the advantages of cryptocurrencies (inflation protection, smart contracts) without forcing society to abandon the existing monetary system.

Thanks for the kind words.

In a sense, it's an alternate form of Proof-of-Stake. As Section 8.4 explains, the conventional wisdom is that Proof-of-Stake's flaw is that it's circular. We've proved that actually it's not circular enough, i.e., the stakes it assigns are different than the stakes in society's existing monetary game.

Proof-of-Balance allows "stakes" (what we call "weight") to be issued in proportion to monetary balances. Once those stakes are in users' hands, the protocol can run using the algorithms designed for PoS, including all of their reward and governance mechanisms.

It turns out that to fully unleash the power of those algorithms, you need a verifiably secure stake-distribution mechanism. That's what we've invented. (It's harder than it sounds, of course...)

Yes, hard to find anything more rewarding than that. Here is the prior HN story on it, if you're interested: https://news.ycombinator.com/item?id=12010760

I wrote about it for the Washington Post, as well: https://www.washingtonpost.com/opinions/how-jeff-sessions-ca...

Sorry about that, here's a dropbox link: https://www.dropbox.com/s/80cpk4ss5ossd4u/KRNC%20Technical%2...

Generally a Show HN has to be something ready for people to use, so I checked with the mods to make sure this paper qualified.

Didn't mean to imply any sort of endorsement beyond that.

Hi HN,

Lead author here. The mods have graciously given me permission to announce some computer science work as a Show HN. It concerns permissionless Byzantine consensus – the notoriously difficult problem of how to securely replicate a state machine in the absence of a reliable identity system, which is the underpinning of Bitcoin and other decentralized ledgers.

By copying the signaling techniques used by animals, my co-author and I have achieved a 40,000x improvement in security and performance over the prior state of the art. This vindicates a prediction made 10 years ago by a Chinese researcher, one of the world's rare dual-PhDs in biology and computer science, who believed that reverse-engineering animal-communication networks could produce a consensus-protocol breakthrough similar to the invention of public-key cryptography.

The parallel between asymmetric encryption and our discovery goes beyond the scale of the advancement. It actually concerns the mechanism that our protocol uses to protect itself from pseudo-spoofing or "Sybil" attacks, in which an entity uses sockpuppets to hijack consensus by casting extra votes. Existing technologies, like Proof-of-Work and Proof-of-Stake, are symmetric in the sense that they require correct agents to "outbid" the adversary by verifiably expending more money or computing power. If the adversary's budget for an attack is greater than the security budget of honest protocol participants, then the entire system collapses.

Our paper introduces the first asymmetric system, Proof-of-Balance. It guarantees that honest protocol participants will remain in control of the consensus protocol, even if their security budget is many times smaller than the adversary's budget for an attack. This verifiable asymmetry yields not only an exponential improvement in security, but also a corollary increase in performance: because the adversary's maximum fraction of total voting power is tightly constrained, transactions can be processed on the open internet using speeds that were previously possible only on permissioned networks.

Asymmetry is nothing new in access control – e.g., a lock increases the security of a house by more than its purchase price, so homeowners aren't forced to "outbid" burglars to keep their families safe. However, it has been ignored in resource-weighted consensus, because the field has been guided by the "handicap principle" – which claims that the reliability of a signal depends on its verifiable cost to the signaler. Bitcoin enthusiasts often expressly invoke this principle to justify the waste inherent in Proof-of-Work, claiming that it is a universal law of nature, which applies with equal force to biology and computer science.

Not so. That is close to what biologists believed in the 1990s, when formal game-theoretic modeling first substantiated the concept of handicap-authenticated signaling. However, subsequent work revealed that it is actually the verifiable cost of faking a signal that determines whether information can be transmitted reliably. If the cost for a dishonest entity to spoof a signal is sufficiently high, then honest agents can transmit reliable signals at zero cost. This is known as cue-authenticated signaling, and it is the key to our protocol, KRNC ("Key Retroactivity Network Consensus").

An intuitive example of the difference between handicap-authenticated signaling and cue-authenticated signaling is how male peacocks and tigers signal their fitness to potential mates. Male peacocks waste resources growing oversized tails, a handicap that proves their fitness based on the amount of self-inflicted punishment they can endure. Male tigers compete with one another to grow as large as possible to gain an edge in lethality, and their size happens to have the added bonus of providing a cue of their fitness.

We adapt the "cue principle" to obtain a novel solution to Goodhart's Law, the adage that a measure ceases to be accurate once it becomes a target. Our rejoinder: if whatever you measure will become a target, measure the thing that is already a target. (The math confirms this.)

For human agents, the universal economic target is money, so that is what Proof-of-Balance uses to assign weights in a consensus protocol. Specifically, it uses mean bank-account balances during a specified window of time in the past – analogous to a "hard fork" of the data in the commercial banking system onto a new cryptographic protocol. Everyone with online banking can unlock their pro rata share of voting power for free. No buying stake, no wasting computing power.

The other major upside to this approach is that it eliminates the need to introduce a new currency, like Bitcoin. Instead, cryptographic weight functions in a similar way to a "symmetallic standard," in which the base money is a meta-resource derived from gold and silver in a specified ratio. In KRNC, base money is a combination of an original fiat unit of account like a U.S. Dollar, plus the corresponding quantity of cryptographic "weight" needed to "back" that dollar.

The difference from the gold standard is that the "backing" isn't entrusted to a Central Bank, which can renege on its word. It's held by the actual users of the money, who transfer both the original dollar and its backing to one another in each transaction. This provides inflation-protection like Bitcoin, but it's added to the world's existing money. No pyramid-scheme like distribution, no risk of technological disruption destroying innocent people's savings.

Formalizing the discoveries in this paper has been, by far, the hardest thing I've ever done. I'm nervous but excited to share the results with the world. I believe they can be used, not just to build faster distributed ledgers, but to protect humanity from the risk of a global monetary crisis. If anyone would like to get in touch, I'll be around to answer questions in the comments, and my email is footnoted on the first page of the paper.

p.s. I'm patenting the technology as part of getting the protocol off the ground, but it's not my goal to be the next Mark Zuckerburg or Larry Ellison. I got involved in this because I freed an innocent man from prison and wanted to see how much more good I could do in the world. If KRNC succeeds on the scale I think it could, I want to use the money for effective altruism and existential-risk reduction. It's the right thing to do.

Thanks for giving me an excuse to discuss Sherdog on HN!

They're the two places I hang out most.

Somehow, I suspect the intersection of their user bases is pretty tiny...

Sherdog has suffered a lot from losing traffic to Reddit. The main forum still has enough users that it superficially seems similar to how it was 10+ years ago, but the subforums are dying.

F12, the grappling forum, used to be one of my favorite places on the internet. It barely has new content anymore. There has, for reasons I don't fully understand, been a mass exodus to r/BJJ.

I can't overstate the loss of community that occurred as a result. On reddit, the fact that usernames aren't prominent and there are no avatars makes it impossible to build "characters" in your mind. People are friendly, but they are perpetual strangers.

Previously, I thought the upvoting and downvoting structure of Reddit was great. But, when applied to a hobby I love, the results were soulless and depressing. Opinions seem to be tailored to the crowd in a way they never were on the F12 forum.

A lot of that may be that Reddit attracts younger people who are newer to the sport. Whatever the reason, it's sad to see a great forum whither away.

Don't take online communities for granted. They seem immortal until they die.

Funny, I posted this today specifically because HN prompted me to.

Not sure what led to my submission of the BBC article from 10 months ago being targeted for a re-post.

But thanks for the link to the earlier discussion!

As an American who spent a year as a visiting researcher inside Russia's Foreign Ministry, I can confirm that these differences in decorum are very real.

They are common knowledge in Russian society, and Russian diplomats are trained to take them into account when emulating Western social graces.

Broadly speaking, in my experience Russians conceive of smiling as something precious, to be shared with friends or family, not wasted on strangers crossing your path. However, if you have occasion to interact with someone (e.g., asking for directions), they may smile, particularly if something about the conversation generates a feeling of closeness. I got a lot of smiles and handshakes in Moscow from strangers when they found out I was studying at MGIMO, since they (incorrectly) assumed it meant I'd snubbed Harvard or Yale.

Howdy Scott,

Is anyone in your field working on the implications of computational complexity for normative ethics? Hume's guillotine relies on the impossibility of any evidence for or against moral facts, but one could stipulate to that while using e.g. Kolmogorov complexity to select moral facts with the highest prior probability to a naive computational oracle.

Not saying it's how I'd necessarily choose my ethics, but if AGI employs algorithmic inference (e.g., approximating Solomonoff induction) for conventional epistemology, the potential may exist for it to extend those algorithms to normative judgments, for better or worse.

These are great questions, but I feel awkward hijacking a thread that isn't really about my experience. The Dateline NBC episode does a great job telling the whole story, including how I got involved. My Washington Post article explains how we persuaded the DA's office.

I am an autodidact, so diving into new subjects is something I naturally enjoy doing. The obsessive focus and intellectual flexibility that helped me in other areas, like computer programming and missile-defense research, worked just as well for forensics and criminal-profiling.

One example that sticks out is how I was able to refute the opinion of the FBI's top criminal profiler by going through and reading the reference text that is the equivalent of the DSM, but for behavioral analysis. He was listed as a co-author, so it would have been easy to assume that he was applying the standards correctly, but I was able to show that his thesis did not match the evidence.

The hardest question to answer is "why this case?" It is as close to a miracle as I've ever experienced ... some force I can't explain caused me to stay up late one night and watch an old TV-special about the murder on my computer, even though I don't watch TV and had zero prior interest in the true-crime genre.

The words of the accused professing his innocence before God were so powerful that, even as an atheist at the time, I felt compelled to look into the case more.

Being used as an instrument to correct a horrible injustice made me reconsider my belief in a higher power.