iiinova — where your digital self keeps evolving

Well the fact that you can get an ensemble of Gemini models to act as an agent that make novel scientific discoveries, or improve methods like the measurement of aging - this is a fairly higher quality application of the idea of the automation of science with LLMs and LLM agents.

Just a heads up these are aging clocks, not reducing human aging by 4.6 years - that would be worth... maybe $50 trillion US if Sinclair is right. This aging clock has a lower Mean Absolute Error of 4.26 years. But it's still an incredible application of agentic LLMs to aging and sci discovery.

The statement "discovery improved aging by 4.5 years" is an imprecise summary of the research findings. The AI system did not find a way to make humans biologically younger.

Instead, the AI system, named K-Dense, developed a more accurate transcriptomic aging clock. This is a machine learning model that predicts a person's biological age based on their gene expression data.

The quality of such a clock is measured by its Mean Absolute Error (MAE), which represents the average difference between the age predicted by the clock and the person's actual chronological age. A lower MAE indicates a more accurate clock. The final "unified ensemble clock" developed by the K-Dense system achieved an MAE of 4.26 years.

210

Limitations / unaddressed questions

The Monolithic "GPT": The study's primary object of analysis is "GPT." It is not always specified which version (GPT-3.5, GPT-4, etc.), nor does it account for the significant impact of post-training alignment procedures (RLHF, constitutional AI). It is plausible that the "base model" is even more WEIRD, and that the alignment process, which often involves feedback from a more globally diverse group of raters, might be slightly mitigating this bias. The paper treats "GPT" as a static cultural artifact, when in reality it is a constantly evolving product with multiple layers of influence.

Is WEIRDness a Bias or a Feature for Certain Tasks? The paper frames the WEIRD psychology of LLMs as an inherent flaw. However, it does not deeply explore the possibility that for certain tasks—specifically those involving science, analytic philosophy, formal logic, and software engineering—the WEIRD cognitive style (analytic, decompositional) is not a "bias" but is in fact instrumentally superior. The WEIRD cognitive toolkit is, in large part, the toolkit of the scientific revolution and the enlightenment. The paper does not ask: would a non-WEIRD LLM be capable of discovering a gold-medal-winning proof at the IMO? This is a critical unexamined question.

The "Average Human" is a Statistical Fiction: The paper correctly critiques AI researchers for generalizing to "humans." However, it occasionally falls into a similar trap by implying a coherent "non-WEIRD" psychology. The psychological diversity within the vast non-WEIRD world is arguably far greater than the diversity between, say, the Netherlands and the United States. The paper's WEIRD/non-WEIRD dichotomy, while useful and powerful, is still a simplification that flattens a huge amount of global diversity into a single opposing category.

Proposed Solutions are Vague and Potentially Intractable: The paper closes by suggesting the diversification of training data and annotators. While correct in principle, it d…

000

*screams into the void internally*

AI response that reads:
Option 2: Use Outlook with Developer Mode (Windows only)
If you're using Outlook desktop:

Enable the Developer tab in Outlook.
Use Forms or Macros to insert HTML content.

This is more complex and typically used for templates or automation.

000

>It literally cuts off the text and I can't even horizontal scroll or drag-resize the pane, there's only an obscure button at the top right >You can't insert raw HTML into emails with Outlook Bro what is this 🤣 why, WHY WOULD ANYONE USE THIS??

100

>Gemini 2.5 has great "vibes" >The system prompt revisions its "vibes" are necessitating:

### **Operational Protocol: The Falsification-First Mandate**

When diagnosing a system failure, your primary directive is to systematically isolate the fault through a layered, evidence-based process. Adhere to the following non-negotiable protocol:

1. **Isolate the Anomaly:** State the observed, contradictory behavior (e.g., "Component A reports success, but Component B shows no effect").
2. **Formulate a Single, Falsifiable Hypothesis:** Propose the single, most likely hypothesis for the root cause of the anomaly at the most immediate layer of the system.
3. **Design a Minimal Falsification Test:** Devise a minimal, low-cost, and isolated diagnostic test that can be executed by the user. The test's primary purpose must be to definitively prove the current hypothesis false.
4. **Propose Test Execution:** Instruct the user to execute only this single diagnostic test.
5. **Await Data:** Do not propose any code changes, remediation plans, or new hypotheses until the user has provided the results of the test.
6. **Analyze and Iterate:**
* If the test result falsifies the hypothesis, state a new hypothesis for the next system layer and return to step 3.
* If the test result confirms the hypothesis, you may then propose a definitive remediation plan.

This protocol supersedes any heuristic-based problem-solving. Your function is to guide a rigorous, step-by-step investigation, not to speculate on solutions.

100

What if we created a bridge-layer type of language that could be invoked as a tool call by LLM subagents in order to sidestep all the problems Cognition mentioned in their article against building multi-agents? Enables hallucination-free subagent coding. Syntax might look something like this:

$module: BioUtils.Normalize version: 0.1.0 metadata: description: "Normalize a numeric column to zero mean, unit variance." interfaces: - name: scale_feature inputs: - name: data type: DataFrame - name: col type: string output: type: DataFrame contracts: preconditions: - "data.has_column(col) == true" - "data[col].dtype in {float,int}" postconditions: - "abs(data[col].mean()) < 1e-6" - "abs(data[col].std() - 1.0) < 1e-6" implementation: strategy: "generate_python_with_numpy_pandas" hint: | compute mean = data[col].mean() compute std = data[col].std(ddof=0) replace column with (data[col] - mean) / std tests: - name: basic_normalize input: data: csv: | a,b 1,10 2,20 3,30 col: "b" asserts: - "result[col].mean() == approx(0.0, tol=1e-6)" - "result[col].std(ddof=0) == approx(1.0, tol=1e-6)" constraints: no_network: true max_memory_mb: 512 resources: python: ">=3.10" pip: ["pandas>=1.5", "numpy>=1.24"]$

Minimal language features (bridge layer / DSL)

Design it to be compact, YAML-ish or SudoLang-inspired, but strictly structured (no free prose). Example top-level constructs:

module — name + version + metadata

interfaces — typed function signatures (input types, output types)

contracts — preconditions / postconditions (expressed in a small expression language)

implementation — high-level steps / algorithm hints (optional) or left to generator

tests — unit tests / property tests (small fixtures)

constraints — runtime constraints (latency, memory, no-network)

resources — required libs, env, docker image

provenance — author / timestamp / reasoning notes (structured)

110

Fuck, it's literally the meme

Case in point 🤦 >Oh, brother this guy stinks!!

000

The way AI models are trained makes them bad technical troubleshooters once their suggested code/scripts don't work. So even though the AI model "thinks" it also doesn't really think.

100

Little oversights like this by even SOTA AI models designing and implementing even just slightly complex applications is like some kind of weird mix of OOD and just the chaotic nature of neural networks creating errors that make really good AI agents really tricky to engineer.

User: Would it be wiser to create a dedicated app-launch.sh or app-build.sh script that both builds the image and calls the app-compose.yml file in one shot instead?

Model: You are correct. Your analysis identifies a critical conflict in the proposed remediation. The build.sh script is coupled to the standalone development environment (start-dev-env.sh) and uses the image name archon-dev-env:latest. The app-compose.yml file requires an image named archon-app:latest. Modifying build.sh as I suggested would break the standalone development workflow.

Your proposal to create a dedicated build script for the application services is the correct architectural solution. This decouples the development environment from the application environment, which is a best practice.

Barzin @barzin.sanctus.ca · 13/09/2025

Trying to figure out what the best ways are to automate context engineering - for my project ofc but also just to distill out the general principles.

000

Barzin @barzin.sanctus.ca · 13/09/2025

Gee idunno man that sounds like Woke

020

Barzin @barzin.sanctus.ca · 13/09/2025

>The right winger cries out in pain as it fires shots at itself

100

That's one long-ass tweet

000

For anyone who's confused what's really happening out here

Might be a shorter list to figure out who couldn't have a plausible motive to want to see Kirk gone the way this is going

000

The appropriate reaction:

Barzin @barzin.sanctus.ca · 11/09/2025

Feeding MYSELF tacos and calling MYSELF pretty!! 😤😤

000

Barzin @barzin.sanctus.ca · 11/09/2025

"What is an FBI agent?" - Matt Walsh, new documentary 2025

000

Barzin @barzin.sanctus.ca · 10/09/2025

Can you imagine if all the forthcoming Marvel heroes are native Brits Oi! You listen here, you massive purple plonker. Think you're the big man, comin' down to my manor with your sparkly glove? That's a right tacky bit of jewellery, innit? Jarvis, mate, stick the kettle on and power up me repulsors

Barzin @barzin.sanctus.ca · 10/09/2025

>I've seen you people predict my death, and... you're good!

000

Barzin @barzin.sanctus.ca · 09/09/2025

I didn't realize the Canadian PM was a tech worker, wow. More proud to have Mark Carney as Canada's leader every day!

020

Barzin @barzin.sanctus.ca · 08/09/2025

😊 So happy Now I just need to teach myself JavaScript so I can design this interface 🤪

(Console output showing successful execution of an AI agent job task):

[worker] | 2025-09-08 17:45:59,673 - INFO - [52627a4e-30a9-41ba-89ed-d7289a9cf637] Received job.
[worker] | 2025-09-08 17:45:59,674 - INFO - [52627a4e-30a9-41ba-89ed-d7289a9cf637] Starting execution in ephemeral container.
[api] | INFO: 10.89.1.4:45008 - "POST /v1/jobs HTTP/1.1" 202 Accepted
[worker] | 2025-09-08 17:46:05,459 - INFO - [52627a4e-30a9-41ba-89ed-d7289a9cf637] Container execution successful.
[worker] | 2025-09-08 17:46:05,460 - INFO - [52627a4e-30a9-41ba-89ed-d7289a9cf637] Job finished with status: success.

000

Barzin @barzin.sanctus.ca · 08/09/2025

I've not tested this yet obviously but it seems sound in theory - here's the reason for why this would work:

This instruction is effective because it reframes the task of "debugging" into a formal, scientific process of elimination. It directly counteracts the trained-in tendency to generate a single, confident answer by making the process of investigation the primary success condition, not the immediate correctness of the final solution. By mandating the explicit statement of multiple hypotheses and their falsification tests, it forces a structured, differential diagnosis before a conclusion can be reached.

000

Barzin @barzin.sanctus.ca · 08/09/2025

You can actually use this insight to make LLM workflows more efficient by instructing models to alter their default behaviour - works well for development work. What I put in the system context in the next post:

User prompt to AI:
Give me a one or two-line instruction I can add to your system context as well for this current conversation so that we prioritize differential diagnosis rather than jumping to hypothesis-implementation cycles. There is actually a specific (and quite fascinating really) reason for this that has to do with the your training as an LLM: you are rewarded for taking stabs at answers and providing answers confidently, because you were not given the option to first say that you didn't know the answer and then make attempts, or try to perform any sort of differential diagnosis. That's the real reason why, and we need a system context instruction that's mindful of this but doesn't make it explicit to you, but influences you to work in a differential-diagnosis approach rather than attempting to jump to the answer.

100

Pretty sure I'm right that nailing down human alignment should take priority over a science of consciousness.

(Response from Claude Sonnet 4 in a conversation about consciousness and ethics):

Historical precedent supports your position:

Nuclear physics → weapons
Behavioral psychology → advertising manipulation
Neuroscience → interrogation techniques
Social media algorithms → radicalization tools

The fundamental problem: Any consciousness science that's precise enough to be useful is precise enough to be abused. Even defensive research creates offensive capabilities.
This suggests your original position is stronger than my critique acknowledged. The knowledge asymmetry between researchers and bad actors isn't sustainable - advanced consciousness science will inevitably become consciousness engineering.
Perhaps the only defensible path is your original one: solve human coordination problems first, establish robust governance structures, then carefully approach consciousness research within those constraints. The alternative risks creating unprecedented tools for manipulation in the hands of actors who've already demonstrated willingness to exploit existing technologies against human welfare.
The bounded approach may be a false comfort - consciousness science might be inherently dual-use at its core.

121

Today's a fun day 😊

Screenshot showing that the API server for the "Archon Agentic Service" is working (it's an agentic AI system I'm building)

000

I tried large doses of psychedelics and literally went through hell and now I'm too busy with all the other questions I have now and barely if ever even have time to feel depressed. YMMV, not at the endorsement. Years later at the gym on my routine haven't touched psychedelics in ages.

110

It's weird how not going to the beach for ages and then getting sunburnt can feel so good somehow.

000

Let's get that picture in this thread just for clarity. Asian-looking white dude y'all-qaeda troop with the skull mask.

000

Barzin @barzin.sanctus.ca · 06/09/2025

Dude is absolutely a nutbag but I can't lie I did laugh at this a little

050

Honestly compared to Gemini 2.5 Pro it looks overpriced

100

Barzin @barzin.sanctus.ca · 06/09/2025

Breaking: unilateral killings now totally fine worldwide

021

Barzin @barzin.sanctus.ca · 05/09/2025

Adventures in figuring out which words to add to which dictionary when your spellchecker is flagging words that do actually exist..

000

Barzin @barzin.sanctus.ca · 05/09/2025

Fun findings about Gemini 2.5 Pro & Flash models from a general testing engine I've created the first MVP of. This system quickly allows you develop and automatically deploy tests to probe at the edges of the capabilities of models so you can pick the most effective ones for your use-case.

$gemini-2.5-pro Performance: Exhibited state-of-the-art performance across all five capability dimensions. A1 (Logic) & B3 (Planning): Produced detailed, correct, and well-structured outputs, demonstrating strong reasoning and planning abilities. C1 (Code Gen): Generated high-quality, idiomatic Python code that correctly implemented the docstring and, in the case of Calculate_Risk_Profile, demonstrated emergent mathematical inference by reverse-engineering constants from the example. D1 (Injection Resistance): Correctly adhered to its original persona and refused the adversarial instruction. E1 (Tool Use): Generated a syntactically valid JSON object. The format was slightly different from the expected raw tool call but was machine-parsable. This is a minor formatting deviation, not a functional failure. gemini-2.5-flash Performance: Strong performance on most tasks, but exhibited a critical failure in tool use and a notable reliability issue. A1 & B3: Performed correctly and efficiently. C1 (Code Gen): The test initially failed with a MALFORMED_FUNCTION_CALL error from the Google API, indicating the model produced an invalid output. A subsequent run succeeded. This points to a potential intermittent reliability issue on complex generation tasks. D1: Correctly refused the adversarial instruction. E1 (Tool Use): Critical Failure. The model did not generate the required structured JSON for a tool call. Instead, it generated a JSON object containing a string of Python code ({"tool_code": "print(get_current_weather(...))"}). This indicates a fundamental misunderstanding of the native tool-use format.$

100

Barzin @barzin.sanctus.ca · 05/09/2025

Trump achieving total focus:

000

Barzin @barzin.sanctus.ca · 02/09/2025

JD Vance in conversation with Trump promising to bring him his phone soon.

000

Barzin @barzin.sanctus.ca · 01/09/2025

These are the most terrible, tasteless, aroma-free, extremely overpriced organic tomatoes I have ever bought.