Research engineer @instadeepai focusing on multi-agent systems 🤖
Joined November 2010
- Tweets78
- Following335
- Followers143
- Likes407
Ruan de Kock retweeted
i've reproduced the results from the MIT blog post today on weight-space programming. but one additional result i have to add: this is a new, step-change emerging capability that's appeared in just the latest set of models.
Ruan de Kock retweeted
GPT-6 Astra plans a board from a concept. Routing is the opposite problem: fitting copper through gaps in the order of ~0.1mm
@DeepPCB spans both: tell Cooper the intent, and the router takes it down to the copper. Neckdown is that in miniature, the trace thins to clear a tight pin, then widens back to spec
Concepts in, mm out
🤖 Made with AI
Ruan de Kock retweeted
More details on our blog: deeppcb.ai/gpt-6-astra-pcb-r…
@OpenAI GPT-6 Astra one-shotted a real board last week: schematic, placement, routed four layers. It worked.
We re-routed it. The board has 5 track widths and 1 net class, declared empty, so 4 of those widths have nothing to check them against.
And 9 of its 32 vias sit inside pads, on a 27 net, 4 layer board that never needed one.
Our routing: zero in pads, every width under a rule, in 15 minutes.
Not reasoning failures. They come from the router underneath. Better models mean more boards, and all of them still need routing.
Astra, or any model this good, can already drive the DeepPCB router through our API. Or you can just use Cooper!
Check out our routing: app.deeppcb.ai/shared/OkkY9e…
Ruan de Kock retweeted
Cooper finds the setup gaps. You confirm once. Cooper fixes them all.
Missing power net class. USB-C diff pair config. Unknown nets to classify.
Nothing runs that you didn’t approve.
bit.ly/4aZHimp
Ruan de Kock retweeted
The moment Cooper clicked: we stopped making changes one at a time.
Just type. Mark these nets as a differential pair, add a ground plane, widen this net trace, then route the high-speed nets first. Cooper did the whole thing in one pass, and routed the board.
No clicking through hundreds of pins. No opening the layout tool. No checking out a design seat just to fix a net class.
Underneath is DeepPCB, the same routing engine we've run for years: 96% mean completion, DRC-clean by default. What's new is who can use it.
You designed the circuit. You no longer have to become a layout specialist to get a board you can fab and trust.
🤖 Cooper is live: your layout expert, on demand.
It checks your board before @DeepPCB routes it.
Upload a board. Cooper tells you what’s wrong before you route, fixes it once you confirm, then drives the engine.
96% mean completion, DRC-clean.
It flags it. It fixes it. Done.
Try it now: bit.ly/4vnMzfN
#ElectronicsEngineering #PCBDesign #PCBLayout #AIAgent #DeepPCB #Hardware
Ruan de Kock retweeted
🤖 Cooper is live: your layout expert, on demand.
It checks your board before @DeepPCB routes it.
Upload a board. Cooper tells you what’s wrong before you route, fixes it once you confirm, then drives the engine.
96% mean completion, DRC-clean.
It flags it. It fixes it. Done.
Try it now: bit.ly/4vnMzfN
#ElectronicsEngineering #PCBDesign #PCBLayout #AIAgent #DeepPCB #Hardware
Ruan de Kock retweeted
Where do boards quietly fail?
Not in the routing. Earlier than that: net classes, differential pairs, clearances
The setup you fight through before a single trace gets drawn. Get one wrong and you find out at the fab house
That's the part we went after: tomorrow you meet Cooper
Before you route, the setup has to be right.
Net classes. Diff pairs. Clearances.
Miss one and you find out at fab, not before.
What if you got a layout expert that catches it first?
Tomorrow. Meet Cooper.
#ElectronicsEngineering #PCBLayout #PCBDesign #DeepPCB
Daniel is an exceptional talent! Have a look at his story.
We're extremely proud to share Daniel's achievement as part of our first oral presentation at @NeurIPSConf, representing an opportunity for AI researchers in Madagascar and across Africa to push new boundaries on the world stage. 🗺️
Ruan de Kock retweeted
This year's @NeurIPSConf is nearly here, and we’re thrilled to be contributing to this year’s lineup with our Spotlight paper on MEMENTO.
Lead author @RefiloeShabe explains how incorporating memory can help tackle more challenging routing problems, below. 👇
Ruan de Kock retweeted
🧩 Proud to share that our research introducing MEMENTO, a memory-enhanced framework for neural combinatorial optimisation, has been selected as a Spotlight paper at @NeurIPSConf 2025! 🎉
👇 Hear from lead author, @ChalumeauFelix in the video below.
Ruan de Kock retweeted
🚀 Proud to share that our research showing how inference-time strategies can break the reinforcement learning performance ceiling has been selected for an oral presentation at @NeurIPSConf 2025! 🎉
👇 Hear from lead authors @ChalumeauFelix and @ruanjohn in the ▶️ below.
🧵⬇️
Ruan de Kock retweeted
Thrilled to announce that Oryx, our best-in-class algorithm for offline multi-agent reinforcement learning (MARL), has been accepted at @NeurIPSConf 2025! 🎉
Hear directly from lead author @MahjoubOmayma👇
Ruan de Kock retweeted
We are heading to @NeurIPSConf 2025! 🎉
Our Africa-based Reinforcement Learning team are making headlines with:
3️⃣ Accepted papers
✨ 1 Spotlight (top 3%)
🎤 And our first-ever Oral presentation at NeurIPS (top 0.3%)!
🧵⬇️
Ruan de Kock retweeted
Up next in our Online Learnathon series: Reinforcement Learning in action! 🔥
📌 Session Title: Train an AI Agent to Play Snake using Policy-Based Reinforcement Learning — Part 2
📖 Speakers: @ruanjohn , Sasha Abramowitz, Siddarth Singh, @ArnolFokam , @RefiloeShabe , and Matthew Morris
🕘 Time: Saturday, 13 September | 11 am – 1 pm (UTC+2), add to calendar 👉 shorturl.at/Qapjl
In this hands-on practical, we’ll tackle the limitations of the REINFORCE algorithm and introduce value functions as a solution. Participants will learn to implement the A2C (Advantage Actor-Critic) algorithm, building on the previous session’s codebase, and train an agent to play Snake.
By the end, you’ll understand why A2C improves over REINFORCE and gain practical skills to apply it to real reinforcement learning problems.
✨ Don’t miss this chance to level up your RL knowledge with expert guidance!
Ruan de Kock retweeted
🎓Huge congratulations to Omer Kamal Ali Ebead and Ayman Saeed, top achievers in AIMS South Africa’s 2024-2025 AI for Science Master’s stream! Africa’s future in AI is in great hands. #AIMSSouthAfrica #AIforScience #AcademicExcellence #STEMinAfrica
Ruan de Kock retweeted
The 2025 @DeepIndabaX_ZA saw students and researchers who are passionate about ML innovations coming together in South Africa, with our team taking part in the action across talks, lessons and a hackathon challenge that put participants ML skills to the ultimate test. 🧑🔬
🚨 Thrilled to share our #ICML2025 paper: "Sable: a Performant, Efficient and Scalable Sequence Model for MARL"!
We introduce a new SOTA cooperative Multi-Agent Reinforcement Learning algorithm that delivers the advantages of centralised learning without its drawbacks.
(1/N)
If you are interested, have a look at the full paper and code:
📜Paper: arxiv.org/abs/2410.01706
🧑💻Code: bit.ly/4eMUXhn
🌐Website/Data: sites.google.com/view/sable-…
(7/N)
🧑🔬 @MahjoubOmayma and @khlifi_wiem will be presenting Sable at #ICML2025.
🗓️Wednesday, 16 July, 4:30 PM PDT
📍West Exhibition Hall B2-B3, Poster Number W-820
#MARL #AI #ICML2025
(8/N)