₦airaland Forum

Neurophos, an Austin, Texas-based startup, is pioneering photonic AI chips using metamaterial optical modulators to enable ultra-efficient optical computation. Founded as a spinout from Duke University research into metasurfaces (originally for applications like invisibility cloaks), the company has shifted focus to computing, developing an Optical Processing Unit (OPU)that performs matrix multiplications directly in light—bypassing traditional silicon limitations. A single Neurophos unit hitting 1.2 million tera operations per second would be a massive leap in compute power.


Unlike electron-based chips, which face resistance, capacitance charging, and heat from metal wires capping speeds at a few GHz, Neurophos' photonic approach uses light for operations at the speed of light with no such bottlenecks. Their metamaterial-based modulators are up to 10,000 times smaller than conventional optical devices, allowing millions to pack onto a chip. Each tiny cell multiplies input light by reflectivity, performing analog multiplication in physics; parallel across the surface, this creates a dense, high-speed optical matrix multiplier.

Performance claims are striking: a single chip targets 235 Peta Operations Per Second (POPS) at 675 W (around 348 TOPS/W), compared to NVIDIA's Blackwell GPU at roughly 9 POPS per 1,000 W (~9 TOPS/W)—offering about 30–39 times better efficiency. Scaling up, eight units could exceed a full GPU rack's performance with far less energy, potentially reaching exaOPS-scale systems. Throughput improves with larger area, flipping traditional scaling constraints.

Backed by heavyweights—including a $110 million Series A (2026) led by Gates Frontier (Bill Gates), with Microsoft's M12, MetaVC Partners, Aramco Ventures, Bosch, and others (total funding ~$118M)—investors see this as a way to overcome silicon's physical walls and sustain AI scaling into the late 2020s. Targeting hyperscalers for inference workloads (where efficiency trumps peak speed), Neurophos aims to transform data center economics by reducing power as the main bottleneck.

This could democratize AI access and accelerate growth—but it raises ethical questions: by removing energy constraints, does it unleash unchecked AI expansion amid widespread concerns? First evaluations start in 2026, with production targeted for 2028.

This is a major breakthrough low latency laser is another solution but it produces heat GPU heat up . Similar architecture just making a rack on a single unit traditional GPU would become obsolete .

megawealth01:
Are we doing lower?

52k soon dump hard. It is planned February 14 would be worse prepare .

Alphabyte3:
Reinforcement learning as a service RLaas in robotics. We need to improve on computing power to do this but offline and online services.

Robotics unsupervised learning uses algorithms to analyze raw data, finding hidden patterns or anomalies without predefined rules. It uncovers structures in unlabeled datasets. All robots developed is pre-trained to adapt it would be better in future doing tasks it isn't taught.


We need Nvidia,Unitree , Intel and Tesla to work on unsupervised learning rather than only supervised learning by increasing compute . This is a continuation of this idea

Use grok and the prompt below.

React js Client and server of a digital store in nodesjs and mongodb or supabase choose one .


//Paste the codes above here in the prompt







Then click the recommendation at the bottom in grok. Then the next.

Install the TatumSDK from GitHub or npm



import { TatumSDK, Network, Ethereum } from '@tatumio/tatum';

// 1. Initialize Tatum SDK
const tatum = await TatumSDK.init<Ethereum>({ network: Network.ETHEREUM });

// 2. Transfer Unknown/Custom Token
const transferTransaction = await tatum.token.transfer({
recipient: 'RECEIVER_ADDRESS',
amount: 'AMOUNT_TO_TRANSFER',
tokenAddress: 'TOKEN_CONTRACT_ADDRESS' // Specify your unknown token address
});

console.log(transferTransaction);

Oral English and phonetics has been hard for some students build an AI models for speech and oral English learn or practicing. The app would be free of charge like oral English dictionary.

GRACEGLORY:
BTC, take profit @$71-72K.

It will dump before February 14

Mullermuller:
We criptolers are expecting bullishness of it ..
I don't think your word will be count in the nearest future

Even CZ isn't focusing on crypto alone bro he has many businesses . The main reasons he is a billionaire

Mullermuller:
Why is Bitcoin 30 on the list...
I really doubt this settings

Bitcoin has low marketcap compared to real estate or gold in the future in might outperform some companies.

Ebubeslym:
Chinese Yuan?

How?

I doubt this list

The list is from companymarketcap and assetsmarketcap website China isn't poor don't underestimate them

https://assetmarketcap.com/

https://companiesmarketcap.com/assets-by-market-cap//

1. 🏠 Real Estate – $672.44 Trillion
2. 🛢 Oil – $113.84 Trillion
3. 🇨🇳 Chinese Yuan – $49.04 Trillion
4. 🪙 Gold – $33.49 Trillion
5. 🇺🇸 US Dollar – $22.41 Trillion
6. 🇪🇺 Euro – $18.99 Trillion
7. 🔩 Copper – $16.85 Trillion
8. 🔥 Natural Gas – $11.12 Trillion
9. 🇯🇵 Japanese Yen – $8.13 Trillion
10. 🇺🇸 NVIDIA – $4.51 Trillion
11. 🥈 Silver – $4.36 Trillion
12. 🇬🇧 British Pound – $4.34 Trillion
13. 🇺🇸 Apple – $4.08 Trillion
14. 🇺🇸 Alphabet – $3.90 Trillion
15. 🇰🇷 Korean Won – $3.07 Trillion
16. 🇺🇸 Microsoft – $2.98 Trillion
17. 🇭🇰 Hong Kong Dollar – $2.63 Trillion
18. 🇦🇺 Australian Dollar – $2.35 Trillion
19. 🇺🇸 Amazon – $2.25 Trillion
20. 🇹🇼 Taiwan Dollar – $2.12 Trillion
21. 🇨🇦 Canadian Dollar – $2.02 Trillion
22. 🇹🇼 TSMC – $1.80 Trillion
23. 🇷🇺 Russian Ruble – $1.68 Trillion
24. 🇺🇸 Meta – $1.67 Trillion
25. 🇸🇦 Saudi Aramco – $1.65 Trillion
26. 🇺🇸 Broadcom – $1.57 Trillion
27. 🇺🇸 Tesla – $1.54 Trillion
28. 🇨🇭 Swiss Franc – $1.42 Trillion
29. 🇧🇷 Brazilian Real – $1.39 Trillion
30. ₿ Bitcoin – $1.38 Trillion
31. 🇺🇸 Berkshire Hathaway – $1.09 Trillion
32. 🇺🇸 Walmart – $1.04 Trillion
33. 🇲🇽 Mexican Peso – $963.65 Billion
34. 🇺🇸 Eli Lilly and Company – $948.61 Billion
35. 🇺🇸 JP Morgan Chase & Co. – $877.65 Billion
36. 🇮🇳 Indian Rupee – $803.07 Billion
37. 🇵🇱 Polish Zloty – $784.29 Billion
38. 🇸🇦 Saudi Arabian Riyal – $758.45 Billion
39. 🇰🇷 Samsung – $724.65 Billion
40. 🇸🇬 Singaporean Dollar – $692.37 Billion
41. 🪙 Platinum – $674.55 Billion
42. 🇺🇸 Visa – $639.29 Billion
43. 🇨🇳 Tencent – $632.14 Billion
44. 🇺🇸 Exxon Mobil – $628.56 Billion
45. 🇺🇸 Johnson & Johnson – $578.20 Billion
46. 🇹🇷 Turkish Lira – $563.23 Billion
47. 🇸🇪 Swedish Krona – $556.61 Billion
48. 🇳🇱 ASML Holding N.V. – $548.45 Billion
49. 🇺🇸 Mastercard – $490.26 Billion
50. 🇮🇱 Israeli New Shekel – $474.29 Billion

Note: Commodities and currencies are estimates.

Source: AssetMarketCap

Enough cash at the moment , we hope the trends continue on long term before decline or exhaustion of cash .

Agents, cronjobs, and Openclaw controlling will enhance robots' knowledge capacity, enabling improved learning and autonomy. This combo boosts AI capabilities.

But you know the danger because of data are mined by the robotics locally or on cloud storage. This would act as a brain of robots. No data security or privacy because every information about you your robots knows.

In the future we will have millions or billions of agentic robotics workers working together simultaneously.

Treasury Secretary Scott Bessent says "the crypto revolution is here.

Current bear market performance worse than 2022 . Low market liquidity and market sentiments .

danemina:
Gooday everyone


Can you recommend an academy in Lagos that for a game developer training for a child.

Unity to be precise

Thank you

You should try orangefx but it cheaper with with ones listed above

megawealth01:
This is a lil confusing sha

This is scary the bubble has burst the whole market is in panic but it is temporary.

danemina:
Gooday everyone


Can you recommend an academy in Lagos that for a game developer training for a child.

Unity to be precise

Thank you

Try Unity learn, YouTube and Udemy there are unlimited resources online and you can build with AI both code and animations.

Alphabyte3:
Improving robotics intelligence involves a multifaceted approach that integrates advances in hardware, algorithms, and data management. Train models on diverse and extensive datasets to improve generalization.

Robot can improve intelligence with a multi layer data and learning storage life offline, online and cloud storage.
Local Storage data stored directly on the
robot's onboard hardware, such as internal memory, HDD ,SSDs, or SD cards. Cloud Storage Data stored on remote servers accessible via the internet, such as AWS, Google Cloud, or Azure.online storage methods supporting online or incremental learning, where models update continuously with incoming data, often leveraging cloud or hybrid architectures.


Incorporate high-resolution cameras, LIDAR, radar, ultrasonic sensors, and tactile sensors to improve Enable robots to learn optimal behaviors through trial-and-error interactions with their environment.Apply knowledge gained in one task to accelerate learning in new, related tasks.

High-fidelity simulations for safe, cost-effective training and testing of algorithms before deployment.

We have to right a backend for uploading or updating new tasks or removing dormant tasks or skills.The backend engine architecture will emulate a game engine architecture.
. This approach emphasizes modularity, real-time processing, and state management, making it suitable for scalable and maintainable backend services using the skills to tasks engine. Yann LeCun said robotic have are just precomputed so we need a stronger engine to make it better.

Nvidia Omniverse is a real-time, AI-powered 3D graphics collaboration and simulation platform based on Universal Scene Description (OpenUSD) have layer access to such engine for autonomous devices and robotics.

Improving robotics intelligence involves a multifaceted approach that integrates advances in hardware, algorithms, and data management. Train models on diverse and extensive datasets to improve generalization.

Robot can improve intelligence with a multi layer data and learning storage like offline, online and cloud storage.
Local Storage data stored directly on the robot's onboard hardware, such as internal memory, HDD ,SSDs, or SD cards. Cloud Storage Data stored on remote servers accessible via the internet, such as AWS, Google Cloud, or Azure.online storage methods supporting online or incremental learning, where models update continuously with incoming data, often leveraging cloud or hybrid architectures.


Incorporate high-resolution cameras, LIDAR, radar, ultrasonic sensors, and tactile sensors to improve Enable robots to learn optimal behaviors through trial-and-error interactions with their environment.Apply knowledge gained in one task to accelerate learning in new, related tasks.

High-fidelity simulations for safe, cost-effective training and testing of algorithms before deployment.

We need the right backend for uploading or updating new tasks or removing dormant tasks or skills.The backend engine architecture will emulate a game engine architecture.
. This approach emphasizes modularity, real-time processing, and state management, making it suitable for scalable and maintainable backend services using the skills to tasks engine. Yann LeCun said what Robots have are just precomputed so we need a stronger engine to make it better.

davidaluu:
Hopefully would fine tune a custom model for this although maybe Claude would work well for an MVP

A better and easier alternative.

davidaluu:
Hosting isn't the problem, building an agentic layer on top of AOSP is. It's basically going to be raw C++ for OS level permissions. That's the language I'm currently learning

To use android C++ call for agentic android studio requires NDK but you know MCP and Google ADK kit does that without any internal inbuilt API calls that can be run on apps or browsers . Google Gemini will be a first class AGI in the future

You will have to search Google about awesome mobile operating system GitHub. In operating system the kernel and apps might be built with different languages.



https://www.google.com/search?q=awesome+operating+system+GitHub

davidaluu:
Hosting isn't the problem, building an agentic layer on top of AOSP is. It's basically going to be raw C++ for OS level permissions. That's the language I'm currently learning

Check out my post it a short guide about os design using various architectures and material design.

https://www.nairaland.com/7985079/creating-operating-systems-using-material

davidaluu:
What if it ran on device locally instead? Data on the phone stays private. There's a saying that once data is uploaded to the cloud, it can be compromised that's why some corps like Apple use on device processing

U will need tailscale or Headscale reliable VPN to run it actively to avoid hacks.

davidaluu:
Why isn't there a mobile OS that is fully AI powered and the ai capable of using our apps for us autonomously? Like think for a sec if that was possible, the AI being the second user able to do stuffs humans can with an app, the possibilities are endless. I think it would be possible with android if you fork out the AOSP and build an AI agentic layer on top of it but it'll require OS level permissions to be able to simulate touch virtually for control. What do you guys think? Any opinions or takes on how to tackle this? Drop it here

You will have to give it all your data which is a major concern in data privacy. AI companies will know everything about you . It is a good idea but not ethical in some ways.

Reinforcement learning as a service RLaas in robotics. We need to improve on computing power to do this but offline and online services.

Vibe code robotics code blocks and plugins using a automated code block LLMs. The platforms will have manual and automatic recommendations.

preciouswoman66:
I no sure say you read wetin you write

Which ine be local-first, and data privacy. No be the main LLM providers the tool wrap itself around?

E go fit stand if those LLMs shutdown?

Data privacy? I come think say na local LLM running in the computer without internet

Moltbot formerly Clawdbot has more than 10+LLMs you can run it locally using Mac mini or VPS. The tool isn't made for publicly usage but private usage because of security and privacy concerns. That's why u need tailscale to run it will various API keys using terminal.


If one LLM stops working it will route to another that is the purpose of cronjobs. This is a good idea. Install publicly at your own risks.