OpenClaws or OpenClaus?

OpenClaw is the Santa Claus of children everywhere. Why, if I am to believe the stories I hear, it is delivering most of what Musk is promising to deliver with an Optimus future!

Here’s what I see happening:

For a year or two now, this whole “agentic AI” thing has been promising that you’ll have at your disposal your own “Chief of Staff”. Such phrases are used by LLMs as part of some sort of fixed point argot feedback with their human users. You’ll see phrases like “lean into” etc. making their way into the human shibboleths to signal to each other that they’re “in the know” about all this AI stuff. There is now emerging a class of people who are spending huge amounts of money per month on subscriptions to the more advanced LLM versions. Combine these two and you have the Smart Set.

Meanwhile, back here on the farm, those of us without access to Codex et al, make do with our puny little non-agentic coding assistants and whatever passes for our multi-decades of experience that might provide something that actually solves the problem of helping individuals deal with their day to day lives in a world where they’re being forced into gig economic status.

For example, as part of my attempt to get a leg up on this, while I’m waiting for GPT5.2 to get through answering a query for code modification (GitHub - jabowery/vibemod: Coding assistant can't produce correctly formed patches to your Python? Use this.) I’ve looked into Upwork as a place where I can bring in enough money to at least increase the number of queries to Claude for basic programming tasks – at which it is pretty good if you can get the spec right.

Although I was aware of this before, it is obvious that Upwork needs some Work. Specifically it needs to put the individual “seller” in a better position to negotiate.

The Chinese markets used to have this tradition (of which I learned from my SO who was stations with her ex-husband in Taiwan while he was doing U2 Mao-flyovers):

Buyer and Seller each write down their bid and ask and then if the bid is higher than the ask, they split the difference and move on to the next transaction with others.

Generalizing this to system that optimizes a marketplace based on my experience with BTC arbitrage, algorithmic trading, business planning, political economy based on rNPV liquidation value property money, etc etc, I came up with something that I could conceivably implement as I wait for GPT5.2 for up to a half-hour to answer a query for small program mods – or wait for DAYS while Claude (whatever – eg 4.6) stops blocking me from using its superior ability to write code to spec.

Let me give you an idea of what OpenClaus can NOT do as “Chief of Staff”:

Keep track of all the constraints on what resources you have and their scheduling in anything remotely resembling a PERT/GANTT calendar, and maintain a big board prioritized by what you personally value given that “in the long run we’re all dead”.

The last one – about “dead” – is particularly relevant as my brother has just retired from being an ER physician and has complained about being stressed out because he can’t keep track of all the little projects that he’s been putting off all his career. That’s what motivated me to start thinking about this as a use of my time while I’m waiting for these increasingly expensive “grad students” to work on my primary objective:

But for some strange reason, “Hume’s Guillotine” isn’t among the shibboleths like “lean into” and “Chief of Staff” with which you can identify members of The Smart Set.

4 Likes

rNPV is Importance
-∂rNPV/∂t is Urgency


Of course this is just one of the dimensions I’ve reduced in the aforementioned “market” where you might need not just to triage but also arbitrage based on the make-buy tradeoff in delegation to unbound agents’ resources. Contractually bound agents’s resources are presumably dispatchable. And both kinds of resources carry the attendant risk adjustment based on their performance history (ie: credit rating whether internally audited or externally provided) – risk that goes into the ‘r’ of the rNPV.

3 Likes

Well hush my mouth! I DO have access to Codex with my basic subscription to GPT but I ended up getting stung financially by it when it was first coming out. The version I tried was one with incremental charges that accumulate the more the agent(s) go round and round palavering – not the flat rate per month one. I honestly don’t recall how I ended up in that situation but my impression was that I couldn’t get the flat rate GPT to provide Codex.

In any event, I’m using it now and it’s better than Claude.

PS: It’s good to see Musk facing a tiny bit of reality here but he’s already falling so far behind I’m not sure he’ll be able to catch up. Grok is really bad and I don’t think it is because xAI hired the wrong people. It wasn’t even that Musk hired the wrong guys to do the hiring per se. The latter is because the same parasites that tell him not to think about abolishing the 16th Amendment in favor of a net asset tax on liquidation are the ones doing the actual hiring of the people who do the hiring.

Musk’s brain parasites are deeply embedded. I suspect the degree of introspection that will be required to dig them out may take with them a substantial portion of Musk’s brain.

I installed Nemotron-3 Nano 30B and challenged it with my standard “anyone home?” IQ test. It didn’t pass but, oh well, what you expect from a model that runs on an NVIDA 3070 with 8GB:

Why bother? Because I can run a sandboxed “OpenClaw” system restricted to my personal computer with “NemoClaw”.

So I installed NemoClaw:

(sentaxbinary) jabowery@jaboweryML:~/devel$ mkdir nemoclaw
(sentaxbinary) jabowery@jaboweryML:~/devel$ cd nemoclaw
(sentaxbinary) jabowery@jaboweryML:~/devel/nemoclaw$ curl -fsSL https://raw.githubusercontent.com/NVIDIA/NemoClaw/main/install.sh | bash
[INFO]  === NemoClaw Installer ===
[INFO]  Node.js found: v25.7.0
[INFO]  Runtime OK: Node.js v25.7.0, npm 11.10.1
[INFO]  Installing NemoClaw from GitHub…
Cloning into '/home/jabowery/.nemoclaw/source'...
remote: Enumerating objects: 205, done.
remote: Counting objects: 100% (205/205), done.
remote: Compressing objects: 100% (183/183), done.
remote: Total 205 (delta 16), reused 103 (delta 8), pack-reused 0 (from 0)
Receiving objects: 100% (205/205), 396.93 KiB | 2.53 MiB/s, done.
Resolving deltas: 100% (16/16), done.
[INFO]  Pre-extracting openclaw@2026.3.11 with system tar (GH-503 workaround)…
[INFO]  openclaw pre-extracted successfully
npm warn deprecated node-domexception@1.0.0: Use your platform's native DOMException instead
npm warn deprecated glob@10.5.0: Old versions of glob are not supported, and contain widely publicized security vulnerabilities, which have been fixed in the current version. Please update. Support for old versions may be purchased (at exorbitant rates) by contacting i@izs.me

added 672 packages, changed 1 package, and audited 674 packages in 1m

133 packages are looking for funding
  run `npm fund` for details

5 vulnerabilities (1 moderate, 3 high, 1 critical)

To address issues that do not require attention, run:
  npm audit fix

To address all issues, run:
  npm audit fix --force

Run `npm audit` for details.

added 168 packages, and audited 169 packages in 5s

55 packages are looking for funding
  run `npm fund` for details

1 high severity vulnerability

To address all issues, run:
  npm audit fix

Run `npm audit` for details.

> nemoclaw@0.1.0 build
> tsc


added 1 package, and audited 3 packages in 1s

found 0 vulnerabilities
[INFO]  Verified: nemoclaw is available at /home/jabowery/mambaforge/envs/sentaxbinary/bin/nemoclaw
[INFO]  Running nemoclaw onboard…
[INFO]  Installer stdin is piped; attaching onboarding to /dev/tty…

  NemoClaw Onboarding
  ===================

  [1/7] Preflight checks
  ──────────────────────────────────────────────────
  ✓ Docker is running
  ✓ Container runtime: docker
  openshell CLI not found. Attempting to install...
  Installing openshell CLI...
[install] Detected Linux (x86_64)
[install] Installing openshell CLI...
[install] Installed openshell to /home/jabowery/.local/bin/openshell (user-local path)
[install] Ensure /home/jabowery/.local/bin is on PATH for future shells.
[install] openshell 0.0.12 installed
  ✓ openshell CLI: openshell 0.0.12
  ✓ Port 8080 available (OpenShell gateway)
  ✓ Port 18789 available (NemoClaw dashboard)
  ✓ NVIDIA GPU detected: 1 GPU(s), 8192 MB VRAM

  [2/7] Starting OpenShell gateway
  ──────────────────────────────────────────────────
  Using pinned OpenShell gateway image: ghcr.io/nvidia/openshell/cluster:0.0.12
✓ Checking Docker
✓ Downloading gateway
✓ Initializing environment
✓ Starting gateway                                                                                             
✓ Gateway ready

  Name: nemoclaw
  Endpoint: https://127.0.0.1:8080

✓ Active gateway set to 'nemoclaw'
  ✓ Gateway is healthy

  [3/7] Creating sandbox
  ──────────────────────────────────────────────────
  Sandbox name (lowercase, numbers, hyphens) [my-assistant]:

The message:
1 high severity vulnerability
is, of course, highly disturbing given the point of this exercise is to discount the risk of OpenClaw violating my privacy. Grok’s advice is not to worry and continue with the “onboarding” and it seems to be providing evidence that there are no exploits of the vulnerability being reported. But it does say that a reasonable, minimum-quagmire, fix is:

Running npm audit fix (without --force) is low-risk — it applies patch-level updates that don’t break semver compatibility.

There are a large number of “gotchas” that arise when using Grok to assist in this but one of the things that I ran into was the fact that one of Musk’s parasites that’s buggering him while laughing at his users defaults sophisticated users asking difficult questions to the least performant mode! This is precisely the wrong thing to do if you’re trying to get coders to start bringing in the bucks they need to afford the higher performing multi-agent coding assistants!

I’m not saying this will fix Musk’s other problems with the basic subscription access to Grok4.6, but fercryinoutloud at least get the basics of onboarding high value customers right!

While developing this idea into a working prototype, I ran across a colleague’s abandoned patent that would, in a better world, have required me to cut him in on my prospective patent - except that his patent was “abandoned”. This reminded me of being impoverished by the ultracentrifugal rocket engien’s patent fees.

The USPTO’s position is that if an inventor runs out of money after spending his capital on the invention that the invention presumably has no utility because capital sources will recognize its commercial potential if it had utility.

My conversation with Gemini is instructive.

Now that I have a few thousand dollars to work with I may give Claude’s one arm another pull in the old multi-armed bandit explore exploit trade-off. However at present I’ve got several “loose ends” projects pretty far along, exploiting codex including Hume’s Guillotine* as well as the aforementioned executive decision support system.

* I am not as impressed by codex’s utility in this challenge but then I’ve been telling people losses compression is the ultimate benchmark.

At this point HeatmapViewport is mostly composition and orchestration. The next meaningful step is to stop
pure refactoring and start moving one actual runtime behavior boundary:

  • either port the current loading overlay semantics into the React host more faithfully
  • or begin replacing one inline-runtime-specific UI subpath with the React host for the generated page mode.

› when using the word “or” in the exclusive sense (ie: a choice is being made that excludes the other) say
** “xor”**

• Understood.

I’ll use xor when the meaning is exclusive choice, and keep or for inclusive or ambiguous cases.

› recall the AGENTS.md directive about partial ordering as applied to decision trees.

The directive is in AGENTS.md.

The key point is:

  • treat roadmap items as a partial-order decision tree, not a rigid serial checklist

Applied to decision trees, that means:

  • identify prerequisites before taking a step
  • recognize independent admissible branches that can proceed in parallel
  • do not assume textual order implies execution order
  • when multiple branches are admissible, choose the one with the best risk-adjusted value
  • only retire a branch when a principled replacement exists
  • if you take a narrower intermediate branch, record the remaining gap in ROADMAP.md

So the rule is not “follow the next listed item.”
It is “follow the currently admissible item with the best leverage, given the prerequisite structure.”

› in light of this, while it is true that choosing one action over another is “exclusive”, that is only
necessarily true in terms of the immediate decision to allocate the opportunity represented by the decision.
it doesn’t necessrily imply a strong path dependency. in other words, some opportunities are not entirely
lost simply because there was a partial order to their realization. they may merely be delayed, pending
availability of resources.

• Correct.

In the partial-order sense, many decisions are exclusive only as immediate resource allocation, not as
permanent branch elimination.

So the stronger model is:

  • a decision often selects the next admissible step
  • it does not necessarily eliminate other admissible steps from the future
  • many alternatives remain live in the frontier and are merely deferred
  • true exclusion exists only when:
    • one branch destroys prerequisites for another
    • one branch commits to an incompatible invariant
    • one branch consumes a non-recoverable opportunity
    • or replacement semantics intentionally retire the older branch

So in roadmap terms, the choice is often not:

  • A xor B forever

It is more often:

  • do A now, while B remains admissible later

That is exactly why AGENTS.md frames the roadmap as a partial-order decision tree rather than a single
irreversible sequence.

A better phrasing for these cases is:

  • priority choice
  • resource allocation choice
  • current frontier choice

not necessarily:

  • exclusive branch choice

So when I describe next steps under this repo’s rules, I should distinguish between:

  • exclusive in this turn
  • mutually incompatible
  • deferred but still admissible

That distinction matters.

› please bear this in mind when stating “either… or…” situations as you just did.

• Understood.

I’ll distinguish among:

  • immediate priority choice
  • deferred-but-still-admissible alternatives
  • truly incompatible xor branches

and I won’t use either … or … when the alternatives are only competing for current attention rather than
being mutually exclusive.

› reword: "The next meaningful step is to stop
pure refactoring and start moving one actual runtime behavior boundary: either port the current loading overlay semantics into the React host more faithfully or begin replacing one inline-runtime-specific UI subpath with the React host for the generated page mode."

• The next meaningful step is to stop pure refactoring and start moving one actual runtime behavior boundary.

Two currently admissible next steps are:

  • port the current loading overlay semantics into the React host more faithfully
  • begin replacing one inline-runtime-specific UI subpath with the React host for the generated page mode

These are competing priorities for the next allocation of effort, not necessarily permanently incompatible
branches.

› good. now which of these represents the highest rNPV?

• Porting the current loading overlay semantics into the React host more faithfully has the higher rNPV.

Why:

  • lower implementation cost
  • lower design uncertainty
  • higher leverage on the main architectural risk: readiness and responsiveness semantics
  • it reduces a named compromise already documented in ROADMAP.md
  • it establishes a cleaner contract before replacing a larger inline-runtime UI subpath

By contrast, replacing an inline-runtime-specific UI subpath now likely has:

  • higher integration risk
  • more hidden coupling
  • weaker local guarantees unless readiness semantics are already trustworthy

So the better next allocation is:

  • first, make React own loading/readiness semantics more faithfully
  • then use that stronger boundary to replace a specific inline-runtime UI subpath

That ordering preserves the second opportunity; it does not retire it.

› very good! proceed.

• I’m taking the higher-rNPV step now: making the React host own loading/readiness semantics more faithfully.
First I’m checking the current scaffold overlay behavior and the existing inline-runtime readiness markers so
I can tighten the React host without inventing a new policy.

59:16 of this podcast features a common misconception about machine learning self-improvement. While it is true the big money wave depends more on automatic software engineering and that Claude has consistently outperformed others, I have been using GPT.

Why?

Like all aging information pioneers I have a lot of irons in the fire and must pick the highest leverage for my remaining few years of adequate cognitive function. As Seymour Cray quipped, in order to make a unique contribution you must be willing to be different. Claude is great at not being different. But difference is what it will take to make breakthroughs. GPT strikes me as more capable of comprehending the paths not taken going back to the mid-60s when Moore’s law first manifest.

2 Likes

ROFLMAO

I must confess to having spent maybe 0.01% of my time playing with multi-agents acting in different roles to do code before I went back to what my intuition told me which is basically what this paper is about.

Me: It is most ironic that the topic of the example is units and the quantity of Urgency is not presented in drNPV/dt units. Getting the units wrong at the level of executive decision support is systemic disaster whereas losing the Mars orbiter is case-in-point disaster. This is very much like confusing income with wealth in political economic decisions.

Codex: Agreed. At executive level, the unit error becomes a governance failure, not merely an engineering failure.