AI for All

1. COMPUTE OWNER 1.2 Fact Asset ai asset
1

Power of AI for Everyone

There is Nothing wrong with AI,
there is Everything wrong with concentration of AI.

Current AI Problems

Deep Learning based Artificial Intelligence requires massive amount of data and compute power, fitting perfectly into the existing Cloud Computing model, substantially worsening the already massive Digital Inequality caused by Cloud Platforms.

1.1. AI deciding who to Kill

Artificial Intelligence (AI) is starting to be used to pick targets instead of just destroying targets. This is no longer about AI killing humans, this is about AI deciding who to kill.

That is AI is being used across the full spectrum: as the police, the jury, the judge and the executioner.

As Microsoft gave ChatGPT a boost in funding it fired all AI ethics staff, the opposite of what is needed to responsible use of AI.

Your Private Cyberspace's AI capabilities are fully controlled by you for more private and hopefully less violent endeavours!

1.2. Not just Coffees at Cafes

How many people at the BrainWash Cafe in San Francisco know their faces were used to help develop AI technologies for foreign militaries ?

1.3. Hallucination

hallucination-leaderboard
hallucination-leaderboard2780×1580 134 KB

source: prompt engineering

Partition AI Solution

Personalised Artificial Intelligence means higher accuracy and ultimate in privacy.

Your Agent is like your own AI personal helper to operate your cyberspace on your behalf. Training your agent like understanding your speech are done on YOUR device for your benefit instead of the cloud's benefit.

As if using your data in influencing your behaviour is not bad enough, now leaking your data could potentially cause you substantial harm.

Current AI developments are mainly based on machine learning, you need to feed it with your data FIRST, protecting your data is more important than ever before.

No matter how advanced an adversary's machine learning capabilities, it still need to get at your data - data that you can start protecting - NOW.

2.1. AI

88.io is about providing AI capability to the masses, so everyone can decide HOW AI impact their individual lives, instead of just been impacted upon en masse BY AI.

You can experience some of 88.io's AI capability with just ONE CLICK.

text
text495×526 324 KB

It can be used to scan most images - the your weight on your bathroom scale, the number plate of the bus you are about to board, receipt for a meal you just had etc.

AI Ownership

As machine learning acceleration of common AI software is made available on more commodity hardware (e.g. Tensorflow inference used to be accelerated only by Google's own Coral hardware, but now we can use OpenVINO on Intel hardware and TensorRT on NVidia hardware), everyday folks are increasing able perform AI operations that traditionally were only feasible in the Cloud on their own hardware.

2

AGI

3

Al Orchestration

4

Scaling

6

AI Problems

7

AI Non-Deterministic

There are a few reasons to why LLMs results are not deterministic even of the "temperature" has been set to 0 (make the model pick the highest-probability token every time).

  1. Floating-point arithmetic:
    The way processors handle calculations with finite precision can lead to different results depending on the order of operations, which can vary slightly between runs.

  2. Parallel processing:
    GPUs run operations in parallel. When partial results are combined, the exact order can vary, causing tiny changes that lead to different results, as explained in this LinkedIn post.

  3. Dynamic batching:
    The number of requests processed alongside yours at any given time can change the batch size. This affects how the model's computations are performed, which can lead to slightly different intermediate results, notes this LinkedIn post.

  4. Amplification over time:
    Because LLMs generate text token by token autoregressively, even a minuscule difference in the probability of the very first token can cascade into a completely different output later in the sequence.

  5. Implementation variations:
    Some implementations may not perfectly achieve temperature zero. They might clamp the temperature to a very small number instead of a true zero or use separate code paths that aren't fully deterministic, as detailed on Medium.

Reference:

8

26.02 World Model

The 26.02 release of our World Model consists of three small vision models from three very different parts of the world - China, France and United States.

3_models
3_models731×427 81.5 KB

Ollama Model Size Month Role
qwen3-vl:4b-instruct-q4_K_M 3.3 GB 2025-11 Primary
ministral-3:3b-instruct-2512-q4_K_M 3.0 GB 2025-12 Secondary
gemma3:4b-it-qat 4.0 GB 2025-05 Secondary

qwen3-vl:4b has the best benchmark scores and will be the primary model controlling the other two models.

3_models_1
3_models_1708×460 55.6 KB

qwen3-vl-4b-instruct

ministral-3-3b-instruct-2512

gemma-3-4b-it

Ollama Limitations

Ollama has a lot of limitations, the situation will improve once we started to add VLLM and LM Studio into the backend LLM server mix.

  1. Ollama does NOT support a lot of vision models, so our vision model options are limited here.

  2. Ollama does NOT support the large context windows of many models e.g. Qwen3-VL-4B has 256K native context size but Ollama only supports 4K.

9

China AI

10

AI Research

A lot of claims, can they deliver ?