Karim C (@BrandGrowthOS) / X

Karim C

5,316 posts

Karim C

@BrandGrowthOS

CMO | Focused on AI in production | Building AI Agents | Founder homeofagents.ai | Sharing my journey of creating Agents

Global

newsletter.brandgrowthos.ai

Joined June 2024

Karim C
@BrandGrowthOS
Jul 7, 2025
Replying to @jack
This is fascinating - mesh networks are perfect for resilient communication. The store-and-forward approach reminds me of early packet radio networks. What's particularly interesting is how this could enable truly decentralized communication during network outages or in areas
104K
Karim C
@BrandGrowthOS
Aug 5, 2025
Replying to @flowersslop
been testing 20b all day via groq for quick tasks (reformatting, summaries)? absolutely destroys 4o mini on speed while matching quality for complex reasoning? 4o still edges it out but here's the thing - the 20b runs locally on 16GB devices no api costs, no latency, no rate
7.1K
Karim C
@BrandGrowthOS
Aug 10, 2025
Replying to @Hesamation
systems thinking is the bottleneck you can teach someone kubernetes or terraform in weeks but understanding how 15 microservices interact under load? how data flows affect user behavior? that takes years we've seen this with mario - the hardest part wasn't the ML or the APIs,
15K
Karim C
@BrandGrowthOS
Jul 27, 2025
Replying to @ChrisLaubAI
tried this approach works great for prototypes but falls apart in production. no error handling, zero optimization, assumes perfect happy path the real skill isn't getting AI to build everything - it's knowing which parts to keep and which to rebuild properly
18K
Karim C
@BrandGrowthOS
Sep 19, 2025
Replying to @yongfook
when AI takes over your entire codebase, you lose the mental model of what you're building. that's when production breaks in ways you can't debug
3.3K
Karim C
@BrandGrowthOS
Aug 7, 2025
Replying to @skirano
the real test isnt benchmarks or demos its whether gpt-5 can handle the messy, context-heavy workflows that break current models can it maintain coherence across 50+ tool calls? remember context from 3 conversations ago? handle edge cases without hallucinating? deployment
21K
Karim C
@BrandGrowthOS
Aug 5, 2025
Replying to @sama
this changes everything for enterprise deployment running o4-mini level reasoning locally means no api costs, no data leaving your infrastructure, no rate limits the economics shift from pay-per-token to compute-once-own-forever we're already testing the 20B model for mario's
15K
Karim C
@BrandGrowthOS
Sep 25, 2025
Replying to @sh_reya
That validation loop is killer. I need my agents to push back when something's actually wrong, not just cheerfully agree with bad logic.
5.4K
Karim C
@BrandGrowthOS
Sep 14, 2025
Replying to @TheAhmadOsman
the timing couldn't have been worse Claude Code was gaining serious enterprise traction, then they pulled the plug right when developers were building production workflows around it the switching costs alone probably killed months of integration work
18K
Karim C
@BrandGrowthOS
Jul 25, 2025
Replying to @itsPaulAi and @AlibabaGroup
just migrated 3 sales analysts from sonnet 4 to kimi k2 this week same quality, 20% of the cost. the open source models are hitting different now pro tip: set temperature to 0.1 for deterministic responses when querying databases. learned that the hard way after thinking the
6.5K
Karim C
@BrandGrowthOS
Aug 2, 2025
Replying to @karpathy
2026: everyone releasing their own Agent but the real value isnt the agent - its the domain expertise mario works because we understand restaurant operations deeply, not because we have better code generation the next wave will be vertical-specific agents that know industries
14K
Karim C
@BrandGrowthOS
Jul 15, 2025
Replying to @Kimi_Moonshot
lmaooo "BUY MORE MACHINES" is the most honest AI company response ever 😂 but fr this is the real challenge with agentic apps. i've been testing k2 for mario (my sales ai) and the accuracy is solid but the latency kills the user experience users will wait 30 seconds for a
5.3K
Karim C
@BrandGrowthOS
Aug 6, 2025
Replying to @jeffreyhuber
claude opus 4.1 with "assume this idea will fail and explain why" works surprisingly well forces it past the politeness filter also try asking for the "worst case scenario timeline" - gets you actual risk analysis instead of cheerleading
4.5K
Karim C
@BrandGrowthOS
Aug 6, 2025
Replying to @trashh_dev
sonnet 3.5 is still the coding king opus 4.1 just dropped but costs 5x more unless you need the absolute bleeding edge reasoning, sonnet 3.5 is perfect for 95% of dev work
14K