Sailesh

Building Prism AI (Stealth)

[Cue Agents @CueAgents]

Hello world!

[Sailesh @sailesh_balu]

@petergyang Ads move more upstream - AI chats and its interfaces

Also as humans spend less time on digital work (as agents take over), their time on entertainments increase (traditional channels would see increase in ad budgets)

[Sailesh @sailesh_balu]

@blessings_soul Same experience. Within 2-3 months of incorporation and gst, i got approached by 20-30 companies offering me services including major banks who even visited my house directly. I just assumed its publicly available

[Sailesh @sailesh_balu]

@danialhasan In

[Sailesh @sailesh_balu]

@MattPRD @Muse Would love an invite

[Sailesh @sailesh_balu]

@businessbarista Great list, but based on this, I dont think there are much AI-native firms now :). But a great checklist for firms.

I would add - an AI transformation team (for Enterprises), to be the brain and manage change across the Org. Otherwise there is no ownership

[Sailesh @sailesh_balu]

@amasad @NotTomBrown Anthropic is scoring points here. Let them do that :)

[Sailesh @sailesh_balu]

@dotey OpenAI and Anthropic have been swapping places since the beginning.

Market share, Fable, Govt blockage, Hugging face incident, and now this.

[Sailesh @sailesh_balu]

I think this might hurt OpenAI more. Possibly,
- Usage shifts to Anthropic models in short term
- Elon goes full-force to improve Grok models over long term
- Elon blocks all infra supply for OpenAI when he has large market share over it

[OpenAI @OpenAI]

We’re ending our partnership with Cursor following its acquisition by SpaceX. Under our proposal, Cursor’s direct access to our models would end on November 12.

We know that the people most affected by this decision are the developers who rely on OpenAI models in Cursor. We care about their experience in this transition and we’re ready to go above and beyond to support them.

https://t.co/OzuCTzUjfX

[Sailesh @sailesh_balu]

@uvkewate Thanks Yuvraj. Had to check our codebase, and we are safe. Keep educating builders!

[Sailesh @sailesh_balu]

@cdolan04 Very important topic (most brands make mistake here). In addition to Chris's clarification, calculation of CAC key:
- Pull out New customer CAC only. Retention costs go in LTV calc
- Include growth overheads (like sales, CRM sub, etc.,)

& pick the right time period for LTV calc

[Sailesh @sailesh_balu]

@kurtinc @Shopify @kurtinc How is this going? If it works, might try it for some of my clients. Also how are you identifying search terms to optimize for?

[Sailesh @sailesh_balu]

@adamguild Owner coming for Toast. We are witnessing how new-gen AI native companies are taking over market from traditional players in SaaS space. Its the best time to build

[Sailesh @sailesh_balu]

@Jake_Langley Remote or on-site? Might have few leads

[Sailesh @sailesh_balu]

@JoeWelstead 100%. There are so many trust problems in eCom space. Seen this in 10+ stores personally. One reason is lack of competition? (e.g., meta, google, shopify). So much opportunity for companies here

[Sailesh @sailesh_balu]

@patloeber Exciting to see Google top voice & image/video gen. Integration into existing google products could be more effective. Lets see

[Sailesh @sailesh_balu]

@ericciarla Useful caveat: keyless removes setup friction, but it does not remove usage limits. For agents, the practical design still needs rate-limit handling and a path to authenticated usage once traffic grows.

[Sailesh @sailesh_balu]

@anirudhbv_ce Well done. Please do more of these

[Sailesh @sailesh_balu]

A very good read of everyone. The author has explained it soo easily. Gold

[ani @anirudhbv_ce]

1/30 Days of Inference ⚙️

Transformers Architecture 📄 (Attention is All You Need)

Transformers killed the RNN by making sequence processing 100% parallelizable. Every foundational model today owes its existence to this architecture. But while the original 2017 paper proposed an encoder-decoder pair, the modern generative AI explosion is driven almost entirely by scaling just the decoder-only portion.

The entire architecture breaks into the core parts:

🔍 Self-Attention (QKV): Queries (what I'm looking for), Keys (what I have), and Values (the actual content). It calculates how much focus every token should put on every other token.

🧠 Multi-Head Attention: Running multiple QKV projections in parallel to capture different contextual relationships simultaneously.

📍 Positional Encoding: Because the model ingests everything at once (unlike RNNs), we have to mathematically inject the sequence order.

⚡ The Hardware Reality: This architecture replaces sequential processing with massive parallel matrix multiplications- perfect for GPUs. But because attention scales quadratically, memory bandwidth quickly becomes the ultimate bottleneck (we'll get into this later!)

(Attached my raw notes breaking down the architecture! 👇)

📚 Highly recommended reads:
• A great technical read by @viplismism on how Attention works under the hood: https://t.co/AuTQT7EbjV
• 3Blue1Brown's Neural Networks playlist for visual intuition: https://t.co/8byrjwyqQW
•@NVIDIAAIDev's blog on the Transformer model:
https://t.co/h4oxEb4gEZ

Tomorrow: We'll look at the latest attention mechanisms that made Transformers more efficient for inference!

#MachineLearning #Transformers #GPU #30DaysOfInference

[Z.ai @Zai_org]

Introducing GLM-5.3-Flash

- Leading capabilities at a highly competitive price
- Natively multimodal with a 1M-token context window
- A 320B-A18B model released under the MIT License
- Previously previewed as Ox Alpha, running entirely on Chinese AI chips

Blog: https://t.co/tzOmB7gdZP

Available now across all official platforms:

Weights: https://t.co/9LRMahY9Wa
API: https://t.co/VcaQnzYmS9
Coding Plan: https://t.co/Nk8Y98HNhU
ZCode: https://t.co/Peepqv4XSx
Chat: https://t.co/WCqWT0qCQb
AutoClaw: https://t.co/aGEG5HqTTb