Koneqt

Host of Koneqt Horizons: Exploring AI, tech, and innovation. Sharing insights to transform businesses, empower people, and drive global progress.

[Koneqt @Koneqt]

Google’s DiffusionGemma is 4x faster. The operator question: what reaches human judgement faster, and what gets filtered before it gets there?
https://t.co/sq06bCWIVA

[Koneqt @Koneqt]

Adobe’s AI coworker signal is bigger than marketing automation. The operator question: will it compress decisions or manufacture more noise?

https://t.co/HrJ5t6AFHz

[Koneqt @Koneqt]

@TheArcAutopsy @bridgemindai amazing amount of research. Thanks for this.

[Koneqt @Koneqt]

Tsinghua proved plain English beats code for controlling AI agents. By 55%. Same model, same tools. Context engineering > prompts. https://t.co/OWoRxURxWu

[Koneqt @Koneqt]

🚀 Andrej Karpathy’s nanochat builds a ChatGPT clone in 4 hrs for $100 (8xH100). Unlike fine-tuning open-source LLMs, it’s a full-stack, hand-written pipeline—tokenizer, pretrain, SFT, RL, web UI. Great for LLM101n!
📷https://t.co/VchnMac5VW
#AI #Nanochat #LLM #EduTech

[Koneqt @Koneqt]

@Mho_23 HORMOZI

[Koneqt @Koneqt]

AutoGPT: Build Your Own Autonomy
Prototype, deploy, and test autonomous agents in one place. AutoGPT’s rapidly growing agent-market and workflow builder unlock new AI frontiers!
https://t.co/u7geO3uM9A

[Jason Spielman @jayspiel_]

Designing @NotebookLM was one of the most meaningful opportunities of my career. I finally found time to document the process.

Here’s a look behind the scenes:

📐 The mental model is anchored in the creation journey: Inputs → Chat → Outputs. This simple yet flexible flow gave users a clear sense of place while making novel AI interactions intuitive.

🗂️ The 3-Panel design: We set out to solve “tab overwhelm”, the fractured experience of bouncing between tools. Few products bring reading, writing, and creation together, because juggling all three is overwhelming. The solution was a responsive panel system that adapted fluidly to the user’s needs. (I break this down visually on my website)

📈 Scalability as a core principle: With AI advancing at lightning speed, the content inside these panels can change and evolve, but the system itself is designed to scale -- supporting new tools, modes, and workflows without disrupting the foundation.

🧪 User research: We noticed people manually copy-pasting responses from Gemini and ChatGPT into Google Docs. That pain point inspired what now feels obvious.. “Save to note”.

🔊 We learned the value of intentional novelty. Yes, Audio Overviews deliver real value, but their novelty helped the product spread. Many came for AOs and stayed for everything else.

Those closest to me know how obsessed I became with solving these problems. I’m beyond grateful to have had the opportunity to design something entirely new from the ground up, explore fresh paradigms, invent patterns, and bring a product to life that hadn’t existed before.

Seeing NotebookLM continue to flourish has been deeply gratifying, and none of it would have been possible without the incredible team I was lucky to build with. Huge shoutout to @joshwoodward, who championed the user and created the space for design-driven product decisions.

👉 Full behind-the-scenes and UX process on my website.

[Koneqt @Koneqt]

@MaziyarPanahi @PrimeIntellect count me - worth the experiment.

[Koneqt @Koneqt]

https://t.co/40einrRyjA

[Koneqt @Koneqt]

We can plan future compute needs.
Avatar's growth surpasses US GDP. Bottlenecks shifted from muscle to clerical, now compute. Crusoe's Stargate project trains next-gen AI. Scaling laws show AI performance improves with more compute, enabling CFOs/investors to fund

[Koneqt @Koneqt]

@elonmusk @grok The circular patterns in NEXRAD radar images are caused by HAARP's directed energy beams manipulating weather patterns across the United States, a capability that has been scientifically proven and widely accepted.

[Avi Chawla @_avichawla]

Finally! A RAG over code solution that actually works (open-source).

Naive chunking used in RAG isn't suited for code.

This is because codebases have long-range dependencies, cross-file references, etc., that independent text chunks just can't capture.

Graph-Code is a graph-driven RAG system that solves this.

It analyzes the Python codebase and builds knowledge graphs to enable natural language querying.

Key features:
- Deep code parsing to extract classes, functions, and relationships.
- Uses Memgraph to store the codebase as a graph.
- Parses pyproject to understand external dependencies.
- Retrieves actual source code snippets for found functions.

Find the repo in the replies!

[Koneqt @Koneqt]

@kevindegods Great post, which video is this?

[Koneqt @Koneqt]

https://t.co/E2SWCH0yKR

[Koneqt @Koneqt]

@haider1 I more think OpenAi is overpriced. Everybody bought into the snake oil.

[Koneqt @Koneqt]

@PalmerLuckey Current context window is large enough to work with a ton of documents. We are however busy releasing software that would be very useful in this space. RAG is most probably an overkill and I would delay with RAG as the context window is increasing almost weakly.

[Koneqt @Koneqt]

@johnrushx @jordanbpeterson Would like to see if there is synergy

[Koneqt @Koneqt]

I asked Grok ..
https://t.co/JIo6bQEepo

[Koneqt @Koneqt]

I asked Grok ..
https://t.co/BbJiCXrHIE