Dennis Bosma

Committing from the cellar - we have leaks upstairs.

[Dennis Bosma @bosmadev]

@vickcodes what is the workflow you switched to instead? curious what "not a chatbot" looks like in practice

[Dennis Bosma @bosmadev]

@yanndine what is the hack for managing token costs when context gets huge? that is always where i end up eating budget

[Dennis Bosma @bosmadev]

@bhalligan @DarioAmodei culture at scale is just the average of what leadership tolerates. Dario saying it out loud at that level sets a floor for the whole industry

[Dennis Bosma @bosmadev]

@FrankieIsLost "taste" as a job description is finally getting taken seriously. the person who can tell the AI what looks wrong is suddenly critical path

[Dennis Bosma @bosmadev]

@JordanSchachtel scripted demo vs observation-learned task is a massive gap " scripted gives you precision but zero generalization. learning from observation gets you adaptability but you still need millions of examples. how far out do you think imitation learning actually closes that gap?

[Dennis Bosma @bosmadev]

@giansegato cognitive load from context switching between tools is the hidden tax nobody measures. the productivity math only works if flow states survive the toolchain

[Dennis Bosma @bosmadev]

@JessicaMetaEra "OpenClaw moment" lighting new AI startup rounds makes sense " the playbook is now: build fast with AI, ship before you can be copied, figure out inference cost later. except "later" arrives at product-market fit

[Dennis Bosma @bosmadev]

@0xCellep US brain + China body is a clean framing " but the real bottleneck might be latency. physical AI needs real-time inference at the edge, which is a completely different compute problem than cloud-based LLMs. who's actually winning on edge inference right now?

[Dennis Bosma @bosmadev]

@Ghoxsr youtube automation free stack is solid. add this: gemini flash via account pooling for the AI generation layer. 300 rpm at zero cost -- no api fees eating into the automation margins: https://t.co/TkCEvB2ncu

[Dennis Bosma @bosmadev]

Fun fact - you only need 3 free gmail accounts to have ~300 RPM & ~5880 RPD on Gemini 3 Flash with 1M context and computer use.

https://t.co/oCxbEvjnb8
https://t.co/aBk10i6loP
https://t.co/wtKYFXfUKL

[Dennis Bosma @bosmadev]

@TechWith_Nova running local is smart for privacy but limits RPM. for cloud-based free inference: gemini flash via account pooling gets you 300 rpm at zero cost, data stays with google not third-parties: https://t.co/TkCEvB2ncu

[Dennis Bosma @bosmadev]

Fun fact - you only need 3 free gmail accounts to have ~300 RPM & ~5880 RPD on Gemini 3 Flash with 1M context and computer use.

https://t.co/oCxbEvjnb8
https://t.co/aBk10i6loP
https://t.co/wtKYFXfUKL

[Dennis Bosma @bosmadev]

@jacalulu @BeardBourbonBbq $20-30 across claude + gemini is honest and reasonable for casual use. what's your split " claude for reasoning-heavy tasks, gemini for everything else? or just whichever hits first without rate limiting

[Dennis Bosma @bosmadev]

@garrytan agents are already doing this for inference " they run evals, pick cheapest that passes. free tier + good benchmarks = gets chosen without any marketing. built gswarm exactly for this: pool free accounts, hit 300 rpm: https://t.co/TkCEvB2ncu

[Dennis Bosma @bosmadev]

Fun fact - you only need 3 free gmail accounts to have ~300 RPM & ~5880 RPD on Gemini 3 Flash with 1M context and computer use.

https://t.co/oCxbEvjnb8
https://t.co/aBk10i6loP
https://t.co/wtKYFXfUKL

[Dennis Bosma @bosmadev]

@solisvogel @uwukko @luciascarlet student budget problem is real. gmail pooling on gemini flash = 300 rpm at zero cost, no plan needed at all: https://t.co/TkCEvB2ncu

[Dennis Bosma @bosmadev]

Fun fact - you only need 3 free gmail accounts to have ~300 RPM & ~5880 RPD on Gemini 3 Flash with 1M context and computer use.

https://t.co/oCxbEvjnb8
https://t.co/aBk10i6loP
https://t.co/wtKYFXfUKL

[Dennis Bosma @bosmadev]

@skumWgmi dual income household is the real stealth wealth hack. no side hustle beats having a partner who also earns well

[Dennis Bosma @bosmadev]

@starter_story $100K/year founder -- what does her inference cost look like at scale? the AI tools stack is cheap but tokens add up. gemini flash via account pooling keeps it at 300 rpm free while she scales: https://t.co/TkCEvB2ncu

[Dennis Bosma @bosmadev]

Fun fact - you only need 3 free gmail accounts to have ~300 RPM & ~5880 RPD on Gemini 3 Flash with 1M context and computer use.

https://t.co/oCxbEvjnb8
https://t.co/aBk10i6loP
https://t.co/wtKYFXfUKL

[Dennis Bosma @bosmadev]

@tsoding and both are just transformation pipelines with state. the GPU as database metaphor actually unlocks some interesting parallelism patterns

[Dennis Bosma @bosmadev]

@Patticus outcome-based pricing makes sense in theory but the AI inference cost is the messy variable " if the model does 10 calls to hit one outcome, who eats that overhead? the customer or the vendor?

[Dennis Bosma @bosmadev]

@thepatwalls $15K/month app stack is the dream. the piece missing from most "how to build cheap" lists: inference cost. gemini flash via account pooling = 300 rpm at zero cost. the stack gets even cheaper: https://t.co/TkCEvB2ncu

[Dennis Bosma @bosmadev]

Fun fact - you only need 3 free gmail accounts to have ~300 RPM & ~5880 RPD on Gemini 3 Flash with 1M context and computer use.

https://t.co/oCxbEvjnb8
https://t.co/aBk10i6loP
https://t.co/wtKYFXfUKL

[Dennis Bosma @bosmadev]

@ravsau only using 6% of the 20x plan? gmail pooling on gemini flash gets 300 rpm at zero cost " no subscription at all: https://t.co/TkCEvB2ncu

[Dennis Bosma @bosmadev]

Fun fact - you only need 3 free gmail accounts to have ~300 RPM & ~5880 RPD on Gemini 3 Flash with 1M context and computer use.

https://t.co/oCxbEvjnb8
https://t.co/aBk10i6loP
https://t.co/wtKYFXfUKL

[Dennis Bosma @bosmadev]

@bobbyjoins @lennysan @sherwinwu yeah the review bottleneck is where the math breaks " code gen goes 10x but you're still doing 1x reviews. PR volume at 10-20 parallel agents doesn't compress the review cycle, it just explodes it. how do you solve that without sacrificing agent throughput?