Entrepreneur and startup lifer. Builder, maker, product designer and marketer. Dad to four kids & a digital product agency called Ventive. See you at #CES!
If your company does any product development, engineering, or manufacturing process improvement, you probably overpaid your taxes from 2022-2024.
Congress quietly fixed this last July. Retroactively.
The catch: the window to amend 2022 starts closing in weeks.
Here’s the full field guide ↓
https://t.co/mM6WuB8OV7
#RDTaxCredit #SmallBusiness #TaxRefund #OBBBA #Section174
Claude Cowork finishing a job gives me flashbacks to 1998 websites on dial‑up. Images loading line by line while you wonder if the computer froze.
Thanks, @AnthropicAI
CPAs: the R&D credit landscape shifted more in 2025 than in the past decade.
Section 174A, retroactive elections, mandatory Form 6765 Section G for 2026, new documentation standards for refund claims.
If you have clients doing R&D, this field guide covers what changed and what to do now:
https://t.co/mM6WuB8OV7
More people need to hear this:
Your anxiety is driven on idleness. You're anxious because you have a goal you're doing nothing about.
Taking action turns anxiety into entertainment. You learn things. You use the mental energy to create rather than stress. You figure things out.
If you feel anxious today. Take some action.
american open source ai just got its biggest win and nobody's talking about it
nvidia dropped nemotron 3 completely open while everyone's distracted by chinese models and closed gpt releases
30 billion parameters total but only activates 3.6 billion at a time so it runs 380 tokens per second which is 2-3x faster than anything comparable
and when i say open i mean actually open not the fake open where you get weights and a good luck note
full model weights plus all training recipes plus pre-training code plus post-training software plus every dataset they legally can redistribute plus the entire rl environment stack they used for agent training
this is the most complete open release in ai history and it's beating qwen3 by +6 points on the intelligence index while running way faster
1 million token context window runs on 24gb consumer ram crushes swe-bench for small models and people are reporting 110 tokens per second locally on their own hardware
not in some cloud datacenter on your actual computer right now today
hybrid mamba-transformer architecture with mixture of experts which means you get transformer reasoning quality at state-space-model speed without the usual tradeoffs
and they're releasing super 100-120b and ultra 400-500b soon with something called latentmoe that makes massive models run faster than physics says they should
while everyone argues about agi timelines nvidia just made the entire ai training playbook public and handed american companies the keys to compete without foreign dependencies