Project SHAME - Sustainable Human Accountability Metrics Engine

[Hugh_Ray_Hyde]

@Hugh_Ray_Hyde

• The DGX Spark currently loads and unloads the models inefficiently. (Unloads and loads the same models between stories).

That has now been fixed (partly/to the best extent).

Now video creation is 4x faster.

4 videos a day instead of 1.

Image Process

I’m using ChatGPT because we think it is slightly cheaper than the API.
(And I have a free month of Plus )

However I can only work in serial.

It seems to me that if you create multiple images at once in separate tabs you get the

You’ve hit the Plus plan limit for image generations requests. You can create more images when the limit resets in * A VERY LONG AND AGONISING TIME *…

message much quicker.

So sticking to slow serial generation seems faster ^^.

API

If I was using the API, I could create images multi-threaded and with much almost no limit… except my manual review time.
(for example you can create 150 images per minute at tier 4).

And if we automated the process then (in theory) we could create 50 stories in a day.

Because we could create 50 images at once.

The overlying process could instead be:

flowchart TB
    A["Choose 50 stories from a list of stories not produced<br>(OSS-120B Generated)"]
    B["Create the JSON for the selected stories<br>(OSS-120B)"]

    subgraph AUTO["Automated Processing"]
        direction TB

        subgraph IMAGE["Image Processing"]
            direction TB
            H["Generate 500 images<br>(gpt-image-2)"]
            C["AI checks the 500 images<br>from a range of PhPU angles<br>(gpt-5.6)"]
            I{"Images approved?"}
            F2["Regenerate failed images"]

            H --> C
            C --> I
            I -->|Yes| F
            I -->|No| F2
            F2 --> H
        end

        subgraph AUDIO["Audio Processing"]
            direction TB
            L["Create audio<br>(Qwen3 TTS Peter Clone)"]
            M["Align audio<br>(Qwen3 Forced Aligner)"]
            D["Check audio<br>(Whisper)"]

            L --> M
            M --> D
        end

        subgraph DESCRIPTION["Description Processing"]
            direction TB
            E["Create an appropriate YouTube description<br>for each story<br>(OSS-120B)"]
        end
    end

    F["Human review of everything"]
    J{"Human review approved?"}
    K{"Which items need fixing?"}
    G["End"]

    A --> B

    B --> H
    B --> L
    B --> E

    D --> F
    E --> F

    F --> J
    J -->|Yes| G
    J -->|No| K

    K -->|Images| H
    K -->|Audio| L
    K -->|Descriptions| E

That is what I got so far.

Happy to hear any feedback!

[phyde1001]

Well we just moved it forward another step locally with MiniMax H3, turned the awesome OpenAI images into video on the local DGX Spark.

Really missing SORA and looks like we can’t legally post the video just yet because we reside in the UK (,EU or South Korea)… but…

The pipeline is there… For SORA’s return.

Looking for further ideas of where we might go with OpenAI products…

Considered turning the videos into high tech cyber security trojans but not sure how that will go down with parents.

or maybe more

Still for now… We have a great local and OpenAI API pipeline for creating multi-lingual fables.

[phyde1001]

OK, now to qualify the fables against the environment my children actually live in…

Walking 6.4 km per day with my two children creates several different forms of value. If that walk replaces a journey the three of us would otherwise make together by car, 6.4 km is almost exactly four miles. Using HMRC’s current 55p-per-mile allowance for the first 10,000 miles as a broad proxy for the cost of using a private vehicle gives around £2.20 per day of avoided transport cost. For climate impact, roughly 2,336 km of avoided travel in a petrol car per year represents approximately 0.5–0.6 tonnes of CO₂e, depending on exactly what emissions are included. I have valued that at roughly £500/tCO₂e, not because £500 is an official carbon price, but as an illustrative cost of achieving equivalent emissions reductions: a recent UK government evaluation of industrial decarbonisation projects found total expenditure equivalent to £521/tCO₂e reduced, compared with a government Green Book carbon value of £313/tCO₂e. That puts the climate component at approximately £0.70–£0.80 per day.

The largest—and most uncertain—component is health. The WHO’s Health Economic Assessment Tool (HEAT) exists specifically to put an economic value on the reduced mortality associated with regular walking and cycling. For walking it uses evidence showing an 11% lower mortality risk at 168 minutes/week, with benefits capped at a 30% reduction at 460 minutes/week; 6.4 km every day is around or above that upper activity level at HEAT’s assumed walking speed. Applying this type of population-level health valuation gives an adult benefit on the order of £4 per day. Combined with the transport and climate components above, that gives a more directly supportable value of roughly £7 per day, before attempting to value the children’s health benefits. My two children clearly receive health benefits as well—the UK Chief Medical Officers recommend that children average at least 60 minutes of physical activity per day—but there is no equivalently defensible simple £ value for those benefits. If I extend the adult figure as an explicitly hypothetical proxy across all three of us, the illustrative total rises to roughly £16 per day, or around £5,500–£6,000 per year. This is not £16 appearing in anyone’s bank account. It is an attempt to express transport, climate and population-health benefits—normally externalised or invisible—in a common unit.

References: HMRC — Business travel mileage rates; DESNZ — UK Government greenhouse-gas conversion factors; DESNZ/Ricardo — Evaluation of the Industrial Energy Transformation Fund; WHO Europe — Health Economic Assessment Tool (HEAT) for walking and cycling; UK Chief Medical Officers — Physical activity guidelines for children and young people.

---

Conventional accounting sees almost none of this. Driving creates measurable expenditure and economic activity; walking largely disappears from the ledger—even where it reduces resource use, emissions and long-term health risk.

So the Engine turns this into a simple daily rule: we walk to the shop to buy fresh food, practise Chinese with the children as we go, and carry the shopping home—the old-fashioned way.

Each day we keep the receipts. And this is where AI becomes useful again…

[phyde1001]

Another quick diversion… This time to my son’s SLM with LLM Fallback. Keeping him busy ^^.

HRH Projects: SLM with LLM Fallback

HRH Projects: SLM with LLM Fallback Recently I combined my earlier Intent Router project with a smaller local-first conversational model I have been experimenting with. Here is a video showcasing the project: [SLM - Small Language Model] The basic idea is: Let the smaller, inspectable system handle requests when it is confident enough, and only fall back to a larger LLM when it actually needs one. Instead of sending every message directly to something like gpt-oss:20b or gpt-o…

[phyde1001]

AI is Poetry?

Dead Poets Society - "Rip it out" scene

[phyde1001]

@phyde1001

Shake the tree

…That’s my strategy… AI will never die…

If you’re still not following… oops.

OK, so I learnt something today that genuinely surprised me:

By 2023, annual global fossil-and-industry CO₂ emissions were roughly 65% higher than in 1993 — when I was 13, the age my son is now.

Roughly two-thirds more CO₂ in a single year than when I was 13.

Why?

Consumption-based CO₂ per person

Region	1993	2023	Change
US	19.9 t	15.8 t	−21%
UK	11.4 t	7.1 t	−38%
EU-27	10.0 t	7.3 t	−27%
China	2.35 t	7.63 t	+225%
India	0.72 t	1.77 t	+147%
High-income countries	12.3 t	11.1 t	−10%
Upper-middle-income countries	2.38 t	5.31 t	+123%
World	~4.07 t	~4.62 t	~+13%

Source: Our World in Data — Per capita consumption-based CO₂ emissions, using Global Carbon Budget data. Values rounded.

Consumption-based accounting assigns CO₂ to where goods and services are ultimately consumed rather than simply where they are produced. The figures here concern fossil-fuel and industrial CO₂, excluding land-use change.

So this does not mean the average person’s CO₂ footprint became 65% larger.

Globally, consumption-based CO₂ per person increased much more modestly — by roughly 13% over this period.

But the world’s population also increased enormously.

Put the two together:

The global arithmetic

	1993	2023	Change
World population	~5.59 billion	~8.1 billion	~+45%
CO₂ per person	~4.07 t	~4.62 t	~+13%
Global fossil/industry CO₂	~22.8 Gt	~37.4 Gt	~+64%

Approximately:

1.45 × 1.13 ≈ 1.64

This is a decomposition of the increase, not separate confirmation of it.

Total emissions are essentially:

population × emissions per person

But it shows something important.

A roughly 45% increase in population, combined with roughly 13% higher emissions per person, produces roughly 64–65% higher total annual emissions.

So when I say ~65% more emissions, I’m not saying individual humans somehow became 65% more carbon-intensive.

There are simply a lot more of us, while average emissions per person have also risen.

And the geographical distribution has changed dramatically.

Per-capita consumption emissions have fallen substantially in places such as the UK, US and EU, while rising strongly in China, India and across the upper-middle-income economies shown above.

So the story isn’t simply:

“The West kept consuming more.”

It is also this:

The world added roughly 2.5 billion people, while average consumption-based CO₂ per person also increased.

And that matters, because other people have rights too.

People have every right to aspire to reliable electricity.

Transport.

Healthcare.

Refrigeration.

Computers.

Education.

A better standard of living for their children.

We cannot spend generations building our own prosperity and then turn around and tell billions of other people:

“Sorry. The carbon budget is gone. Stay poor.”

That isn’t a serious moral answer.

But recognising other people’s right to development doesn’t make the additional CO₂ disappear either.

And I think that is the actual problem:

How do billions more people exercise a legitimate right to development while humanity reduces the total environmental cost of all of us living well?

Not blame.

Not East versus West.

Not an accounting trick.

More people.

Legitimate aspirations.

~65% more annual emissions.

One atmosphere.

I used ChatGPT to help interrogate the figures and check the arithmetic because I had limited time. The underlying datasets are credited above.

AI didn’t create the numbers.

It helped me shake the tree.

Corrections welcome.

I’m deliberately showing the arithmetic, definitions and sources because I’m trying to account for my own claims, not just throw another number into the discussion.

[phyde1001]

The post above is the accounting.

This is why I’m doing it.

We are extraordinarily good at accounting for money. And money buys labour, expertise, computation — access to intelligence.

But intelligence optimises against whatever we choose to account for.

If the ledger rewards growth, profit or convenience while some physical costs sit outside its boundary, then we can become more intelligent, more efficient and richer — while still making the physical situation worse.

Clean technology shows that we are getting smarter. It is already preventing billions of tonnes of additional CO₂.

And yet the actual global emissions number remains around record levels.

So I’m trying to work out where bad intelligence — artificial or otherwise — enters the system, and what false accounting allows it to keep producing an outcome none of us would consciously choose for our children.

Not by calling out individuals per se, but by identifying exactly where the boundaries are, what falls outside them, and trying to keep within those boundaries myself.

Money can be repriced.

Accounts can be rewritten.

Physics cannot.

Because somewhere, the numbers we optimise for have become detached from the numbers that actually matter.

[phyde1001]

Why is this relevant here?

Maybe this is the eval.

Not: How intelligent is the system?

But:

Can increasing intelligence improve human lives while protecting and restoring the ecosystems we depend on, and reducing the physical cost of doing so?

If capability, wealth and efficiency keep rising while total emissions remain near record levels, then intelligence may be improving — but the system-level eval is still failing.

Benchmark the outcome, not the cleverness.

The cleverness isn’t the hallucination.
Mistaking cleverness for intelligence is.