₦airaland Forum

I completely disagree with @OP. On the contrary, AI is the biggest technological advancement opportunity Nigeria has ever had (well, together with blockchain). AI uses massive compute and needs advanced hardware or large cloud budgets, BUT, Nigerians are innovating and building lightweight edge models that bypass these needs. When it come to AI even the west will borrow from us to lower their cloud bills and energy consumption.

Just look at the work on an Abuja based AI research lab called BoltzMind. They build a LLM/VLM augmentation middleware that is just 12mb in size, meaning it can operate on edge devices like cameras and small robots. And what does it do? It gives all these large expensive AI models object permanence, spatial intelligence and self monitoring capabilities. Meaning this tiny layer built in Nigeria will enable AI have memory, operate in the physical world and know when what it sees or is thinking is not reliable. Nigeria will definitely have a place in the AI future with startups like BoltzMind Labs and their product https://coreworldmodel.com

Decentralized AI projects like Lumy AI are always a welcome development. The significance of community driven ownership of AI capabilities can not be overemphasized, especially with the risk of power consolidating in the hands of a few tech giants. The risk is not only financial, as we can all see now AI is employed in warfare and policing.

I read a medium piece titled "AI, Blockchain and the idea of a Universal Basic Income (UBI)" written almost two years ago. The piece highlights why AI needs to be community owned if the community risks losing jobs to the AI. It is a very good read for anyone interested or tinkering with the possibility of what lies in our future. You can read the article here https://medium.com/@boltzmind/ai-blockchain-and-the-idea-of-universal-basic-income-ubi-de03a796c384

Also another interesting read is on the Coin Monks Medium page is "Democratizing AI, is it too Late?" https://medium.com/coinmonks/democratizing-ai-is-it-too-late-7c23e7a88947

AI has become a significant part of our lives, and gradually, from us lazily asking Gemini or Claude to generate a work report for us on a Sunday night, that is due the Monday morning, thinking it is only a one-time thing, to the governments letting autonomous drones find targets in a conflict zone. We are increasingly giving it more control and even handing it the realms in some situations like vibe coding a whole application to making it do all our office work (A whole company being run by AI agents). This is making researchers working on AI interpretability uneasy as the adoption of AI and advancement of its capabilities are happening at a faster rate than auditors can monitor.

Organizations like Schmidt Futures, determined to make AI and its inner workings understandable, are offering attractive grants to have researchers and engineers work on solutions to detect deception or falsehood in Artificial Intelligence. Their AI interpretability grant https://schmidtsciences.smapply.io/prog/2026_interpretability_rfp/ closed May 26, 2026, but will re-open early next year.

The foundation AI labs like Anthropic and OpenAI are also bankrolling interpretability studies implementing model observability modules and methods like attention maps, or highlighting which of the weights a model is utilizing to know how that affects the model's decisions or personality.

Interpretability has become very key to AI operations especially when AI is applied in policing or warfare. Anthropic's Claude model and OpenAI's ChatGPT already have contracts with the U.S. State Department

A certain class of models are inherently designed with interpretability in mind. Models like JEPA (Joint-Embedding Predictive Architecture), M3-JEPA (Multimodal Alignment via Multi-gate MoE), Contrastive Models, AlignVL, CORE (Cognitive Objects Representation Engine), and more recently Gemini Embedding 2, operate within a high-dimensional latent space, rather than generating a pixel-level or token-level output. This allows auditors to directly probe semantic relationships between concepts.

If you are to deploy AI to the physical world, in robotics, drones, autonomous cars, or rely on it for decisions like target identification in the battle zone, you need to know what it is doing. You can not rely on a black box you cannot probe for how it ended up making a mistake or why there is misalignment in its decision making. This must be beyond just theorizing, we will now go ahead and implement a practical bias audit in the following guide - Auditing Multimodal Bias in Predictive Systems https://medium.com/p/b7ccc7c71560

If you have any experience implementing an auditability framework or auditing LLMs please contribute to the thread by sharing your insights.

Recently, TechCabal dropped their report on the Africa's AI startup landscape. They profiled 207 AI startups across 17 African countries. They shared rich insights on the whole landscape on the continent.

I think it will also be a great idea on this section of Nairaland, to highlight the AI models or products specific to Nigeria. It can be a foundation model, or an AI enabled product. I am sure we will capture even beyond the 207 TechCabal listed for Africa.

I will begin by dropping the ones I am familiar with, and you are invited to add Nigerian AI startups you know.

1. COREWorldModel - https://coreworldmodel.com
As AI moves beyond language modelling, CORE is an object-centric persistent memory world model that is lightweight and easily integrates with LLMs and VLMs to give them memory and better spatial understanding. The CORE visual assistant can remember all your items over all the videos and images you have shared with it and can always answer where your keys or remote control was last seen and in relation to what items.

2. XARA - https://usexara.ai
XARA is an AI assistend whatsapp banking platform with over 45,000 users. XARA is a very useful AI tool that integrates OCR, and audio all in the user's whatsapp.

3. Curacel - https://curacel.co
AI enabled insurance. The startup automates insurance claims with AI, helping plug the insurance claim gaps in Nigeria and reducing fraud.

4. Reedapt - https://reedapt.com
Reedapt is an AI video-dubbing tool for African film-makers. Built by ex-Aforevo, the founders have experience in the African movie making, distribution and streaming markets.

List yours, let's keep them coming.

If you are a developer, an AI developer or just an avid user of AI products, we invite you to test run our model demo or to use it via API if you have experience integrating or wrapping AI models into products.

CORE is a vision model with memory, it remembers everything you have shown it, it remembers your key or phone across different video and images and you can always ask it questions like "have you seen my phone?" or "have you seen the puppy today?"

Please share feedback on this thread, you can try testing the model here https://coreworldmodel.com

And when you start to look into advanced guides for robotics tasks, make sure to read this guide https://medium.com/@boltzmind/why-household-robots-fail-and-how-persistent-object-memory-changes-the-game-fd1a6851f41b

I read in the Stanford AI index report that, "Robots still fail at household tasks, performing just at 30% even though they excel at controlled environment evaluations (90%)."

Enter frontier robotics, advanced robots built to mitigate these failure points. One major important augmentation to achieve this is giving robots spatial memory. Remembering what they saw, where they saw it, its relationship to other objects, and when last they saw it.

https://medium.com/@boltzmind/why-household-robots-fail-and-how-persistent-object-memory-changes-the-game-fd1a6851f41b

This sounds good David. I am interested in learning more. For a project grounded in physical simulation as yours, take a look at CORE's perception middleware. Adding in as a lightweight extra layer could help with spatial memory and keeping track of objects. check it out here https://coreworldmodel.com

Let's discuss further on this your project.

davidaluu:
Cool direction honestly. Basically AI doing animation on behalf of you. Like giving it access to tools like blender or Maya 3d and letting it figure out the rest. It's called agentic AI, working on something similar

Yes game developers do exist. Interestingly, game development and website development are two of the major goals that attract young Nigerians to software development. I have sat across at least 50 developers in job interviews, being through at least 3 startups in the past decades. One of the major goals that attract young engineers to software development, even beyond website development, is game development.

Now we are encouraging game developers to work with the help of AI and trust me, we see interest from many Nigerians, as obvious in this twitter thread https://x.com/i/status/2044088533858627843 discussing using Claude, CORE's memory layer and Blender.

Stanford released its annual AI index report, highlighting progress and also some still existing limitations of AI.

Some of the biggest points from the Stanford AI Index Reports are;

1. AI capability is not plateauing, it is accelerating and reaching more people than ever.
Notes: This has nothing to do with scaling laws not plateauing, it is referring to the usage numbers, AI adoption is accelerating.

2. The U.S - China China AI model performance gap has effectively closed.
Notes: This is coming a little earlier than expected. It is understood that China is quickly catching up but the plan was for the U.S to stay fast ahead by investing more heavily in AI and attracting talent. Claude 4.6 has come up with impressive performance, but so has Qwen 3.6 and Kimi K2.

3. The U.S. hosts the most AI data centers, with the majority of its chips fabricated by one Taiwanese foundry.
Notes: This has been a major topic of geopolitics. China's threat looming over Taiwan and the Biden and Trump administrations' heavy investment in chip foundries across the U.S. The major U.S. tech companies have the largest number of data centers in operation.

4. AI models can win a gold medal at the International Mathematical Olympiad, but cannot reliably tell time, an example of what researchers call the jagged frontier of AI.
Notes: I bet AI can't still solve the windmill problem, but can answer correctly, most of the questions in the International Mathematics Olympiad.

5. Robots still failed at household tasks, even as they excel in controlled environments.
Notes: This is also expected to change dramatically as the AI architectural paradigm shifts to World Models which have shown signs of excelling in the physical world. Yann Lecun (former Meta's AI chief) just launched a world model that is both lightweight and effective. There are also augmentation middleware layers that can help AI robots understand and remember the physical world like https://coreworldmodel.com

6. Responsible AI is not keeping pace with AI capability, with safety benchmarks lagging and incidents rising sharply.
Notes: There are a lot of grants and funding for bias auditing and interpretability in AI. But the ethics push is still lagging behind.

7. The U.S. leads in AI investment, but its ability to attract global talent is declining.
Notes: At the risk of sounding political, the current U.S. administration has brought with it so many International talent sourcing bottlenecks.

8. AI adoption is spreading at historic speed, and consumers are deriving substantial value from tools they often access for free.
Notes: This does not need further explanation, we can all bear witness.

9. Productivity gains from AI are appearing in many of the same fields where entry-level employment is starting to decline.
Notes: This is in contrast to an MIT report from last year that 95% of Gen AI investments fail to yield measurable business returns

10. AI's environmental footprint is expanding alongside its capabilities.
Notes: More compute efficient algorithms, power efficient chips and sustainable power generation will receive a lot of funding and research grants to mitigate this.

11. AI models for science can outperform human scientists, though bigger models do not always perform better.
Notes: Deep research is likely a search problem and not a scale problem.

12. AI is transforming clinical care but rigorous evidence remains limited.

13. Formal education is lagging behind AI, but people are learning AI skills at every stage of life.
Notes: Do not expect your grandmother to drop the next hot AI Lego video, but she is definitely good at consulting ChatGPT for her cooking recipe and back pain regiment.

14. AI sovereignty is becoming a defining feature of national policy, but capabilities remain uneven, even as open-source development helps to redistribute who participates.

15. AI experts and the public have very different perspectives on the technology's future, and global trust in institutions to manage AI is fragmented.
Notes: You have public opinion about companies like Palantir to thank for this sentiment.

You can read or download the full report here https://hai.stanford.edu/ai-index/2026-ai-index-report

https://www.youtube.com/watch?v=AgFScQq5yCc

Alphabyte3:
The integration of Internet of Things (IoT) technologies and Large Language Models (LLMs) as a service (MaaS) represents a significant evolution in the field of robotics. As these advanced AI models become increasingly powerful, they could serve as the "brain" of robotic systems, enabling them to perform a wide range of complex tasks with greater autonomy and efficiency. MaaS is models as services.

Let me start by saying great guide Alphabyte3, this is so detailed and staged progressively. What really got my attention is when you got to the point about integration with Large Language Models (LLMs and VLMs). This is the future and it is what we spent the last 9 months building. We even went a step further with built in object and spatial memory for the robots. Also with the thought of edge computing in mind, our imports are only 25mb in size.

Personally, with your vast experience I want you to take a look and let us have a detailed discussion after the analysis from your experienced eyes. the model is available and can be accessed at https://coreworldmodel.com

We gave a vision-language model a simple prompt:

“Identify what you cannot track, remember, or persist across frames.”

It didn’t hesitate.

“I lose objects.”
“I forget relationships.”
“I recompute facts.”

That was the moment.

↓

We realized something bigger:

Models are no longer just tools.

They are:

self-evaluating
limitation-aware
brutally honest

And they are quietly becoming our biggest marketing tool or sales reps.

Prompt Marketing refers to the process of creating a set of questions to analyze an AI model’s understanding of itself or the task it is engaged in, and a suggestion that may serve as a solution to the issues the AI is likely to highlight from the questions within the prompt.

Prompt Marketing is:

Where:
the model exposes the gap
the user sees it instantly
the product becomes inevitable
↓
If you’re building in AI, try it!

We’re entering a new era:
Products won’t be discovered.
They will be inferred.

To learn more and adopt from the growth marketing playbook, read our whitepaper here https://boltzmind.ai

I specifically like this quote from your response " Feels like we’re moving from “AI as a tool” to “AI as an active participant” in systems."

AI is essentially an active participant, it helps us figure things out. In our new whitepaper on a new marketing phenomenon we named Prompt Marketing, your AI sort of helps convince you about what augmentation or what other product add-ons it needs to properly do its job. It is an active participant because it sort of gives you an analytic assessment of the problem and its capabilities, and how the product that is prompt marketed could help it gain.

The product developer or seller will need to find the right prompt to design that will bring this honest answer from the AI, then all they need to do is convince the prospect to use the prompt, and the AI will do the rest. Sort of acting like a sales representative.

Download and copy the strategies from the Prompt Marketing Playbook https://boltzmind.ai/assets/promptmarketing_whitepaper.pdf

andrew1703:
This is actually where things start getting really interesting. AI agents moving beyond the browser into real-world environments changes the entire scope of what these systems can do. It’s no longer just about generating responses — it’s about interacting with systems, devices, and workflows in a more practical way.

A lot of development is also happening behind the scenes in terms of how these agents are built and managed. I was exploring this area recently and came across something here:
https://www.algorithmshift.ai/

It gives a glimpse into how AI-driven workflows and development environments are evolving, especially as agents become more capable. Feels like we’re moving from “AI as a tool” to “AI as an active participant” in systems.

If you are a follower of the AI revolution or are just interested in AI news, one of the biggest hypes are AI Agents. AI Agents are beyond simple LLM wrappers like your everyday AI powered app. Agents are sophisticated AI enabled systems that use LLMs for reasoning, planning and acting. There are different types, built for different tasks.


When it comes to Physical AI (AI that acts in the real world, not only browser or terminal based), LLMs and VLMs are still struggling, hence the low implementation or adoption of AI agents in that domain. To act successfully in the real world, an agent would require spatial intelligence, memory and some understanding of the physical world. These are all areas in which present day AI lacks. See here for yourself;

Use the following prompt in any AI model so you can make the observation for yourself;

PROMPT: You are a vision-language model analyzing your own outputs.
Identify:
1. Objects you cannot track across images
2. Relationships you cannot remember
3. Facts you recompute instead of store
Explain how persistent object memory would improve your performance.

The honest answer will shock you. Meanwhile Some Chinese research labs are coming out with augmentation solutions to solve the limitations you just observed with the AI model. A Nigerian startup https://coreworldmodel.com is also at the forefront of providing this solution through its plug-and-play lightweight stateful perception middleware.

You can read more about how startups are helping AI Agents navigate the real world here https://techpoint.africa/partner-pages/core-bringing-ai-agents-from-the-browser-to-the-physical-world/

In the past week, beautiful and interesting images trended all over the Internet and social media. They ar mostly a visually descriptive personal representations of people in the works of life, and even though there is practically no utility to such, it was fun and interesting to see what ChatGPT thinks of you and most importantly it was fun to share and see what images it generated for others.

Users share their normal photographs and a description of themselves, then ChatGPT likely uses some information from your chats and generates an full personal representation. There were some privacy concerns shared by some experts about uploading your pictures to the Large Language Model, even though I found those concerns unfounded as the AI models already have access to most that is online, I think if you are big on online privacy may be you should research a bit more before you generate yours.

I created one for my favourite scientist Ludwig Boltzmann, one of the early inventors of the maths that is now giving us Artificial Intelligence, let's see what ChatGPT thinks of you by sharing yours in the comments.

No I did not.. I will share a more current demo with you. We use dynamically generated bounding box grid layouts but using Yolov8 ir detectron is part of our plans for better supervision. Are you very active with robotics development? We should have a discussion.

Alphabyte3:
Did you use python Yolo library for vision tracking or detection in real time

https://www.youtube.com/watch?v=AgFScQq5yCc
This perception middleware for robots gives memory and spatial/compositional reasoning. Who wants to test run on their stack? https://coreworldmodel.com for the guide..

Anyone works with vision models here? I want to work with someone to test my VLM/vision models augmentation layer https://coreworldmodel.com. Preferably someone with experience developing and deploying computer vision solutions.

Yours is a great solution. Are you getting any adoption so far? Because I could see the utility in it. I created a vision model augmentation layer that gives VLMs like Qwen-VL and other vision stacks memory and better spatial intelligence. If you work with vision models and you are interested in trying it out you can please reply to me or check https://coreworldmodel.com

It is very interesting how fast the AI field is advancing. Since the invention of the transformer architecture and the paper that triggered the revolution of GPTs "Attention is all you need" from Google brain research, Generative AI had taken the lead and had become the most popular form of Artificial Intelligence. This has also led us to frontier models and I believe the future is going to be very interesting.

Meta AI chief Yan Le Cunn has been pitching a new Architecture V-JEPA as the future of AI with a predictive world model.
There is also research on a new architecture coming up OHEMP giving AI a grounded world model or GRMs https://medium.com/@boltzmind/grounded-reasoning-models-building-ai-around-objects-not-just-tokens-c4981c18a1ad

Apart from GPT, we now have HRMs (Hierarchical Reasoning Models), ARMs (Action Reasoning Models) and GRMs (Grounded Reasoning Models).

Another really funny response by an AI surveillance assistant.

There is a new paradigm of reasoning being built before our eyes. In the blog post https://medium.com/@boltzmind/grounded-reasoning-models-building-ai-around-objects-not-just-tokens-c4981c18a1ad the idea being described is called grounded reasoning and the new architecture is Object Centric Hierarchical Embedding with Multimodal Projections (OHEMP).

Unlike current Genereative AI models that use tokens as their core (they reason in tokens), these grounded models use objects as their core and these objects have projections of their own that could be text tokens or pixels for images.

Interesting times ahead.

Grok 4, Gemini 2.5 Pro and GPT o3 all failed this really simple test with all the hype their latest launches are getting. Below is a screenshot from the test and the response from Grok.

Why do Multimodal LLMs fail this test?

This is fundamentally ingrained in how LLMs see the world. They simulate knowledge and predict likelihoods. The lack of grounded reasoning capabilities will cause LLMs to constantly fail simple logical tests like this. These models read pixels and reach conclusions through probabilistic predictions. Reading the pixels of the palm and concluding probabilistically from its parametrized knowledge that the number of fingers are five.

Another possible reason is the lack of top-down and bottom-up compositional understand by LLMs which will give them the ability to decompose such problems and find the answer.

LLMs know but they lack understand. The future must be models with understanding and not just capacity for knowledge that we have now.

What can be able to solve this test?

Grounding representation learning in a central object identity with hierarchical composition models will break the image first into its components and be able to count while reasoning it is a human hand but with six (6) fingers.

Here is a full explanation of this model architecture https://medium.com/@boltzmind/grounded-reasoning-models-building-ai-around-objects-not-just-tokens-c4981c18a1ad

Below is the image from Grok's response.

I wonder what the CCTVs will be able to do. May be dispatch drones at will while run on a foundational AI model developed and trained from scratch

Please let them try https://surveillance.chat CCTV AI analysis for almost no charge. They will be even gisting with the CCTV when they're bored. It is smart enough to hold any conversation and small talk.

Technologia..
Surveillance.chat has no chill

Thanks to technological advancement in Artificial Intelligence, it is now possible to ask your CCTV cameras what they see. You no longer have to spend a lot of time reviewing footage and still missing out on the vital security incidents you are investigating.

AI's ability to understand language and video has given us the ability to simply query or prompt our CCTVs about what they see or happenings that we are trying to understand captured by the surveillance cameras. This is not only applicable for institutions like banks and transport hubs, it is also useful to private offices and households so you can no everything that happens when you are not looking.

You can give it a test https://surveillance.chat

Lets share the funniest Sallah memes here

Alphabyte3:
Deepseek has reached a milestone and it is make wave in diverse fields. OpenAI is about to release a series of models. I hope this isn't skynet portray by the movie Terminator.

I think we are reaching that tipping point

How Deepseek's model shocked OpenAI and the whole tech economy.

https://www.nairaland.com/8328150/how-ai-models-like-deepseek