Agi Man

tech senior, playing (programming, architecting SW systems, ...) with computers for 40+ years and still don't have enough ;-)

[Agi Man @AgiMan_WCL]

@MiaAI_lab They are waiting on Anthropic IPO - to drop new models a day before IPO and pop USA AI bubble ? 😉

[Agi Man @AgiMan_WCL]

@TokenGremlin They trained models on other companies know how, now they will take over their bussines, make sure that nobody can distill their models and that nobody can use they models to compete with them. If nobody stops them (pop the bubble), they will eat the economy ... or not😉

[Agi Man @AgiMan_WCL]

@aleabitoreddit The pool of capital of pool of something else? 😉 This looks more like private capital wanting to offload their s..t to bigger suckers (definition of trade by fintech). Will see how this will end ...

[Agi Man @AgiMan_WCL]

@Da7_Tech What books, what better data? RL environments/pipelines are driving model improvements in last year ...

[Agi Man @AgiMan_WCL]

@MiaAI_lab I would wait until the day before Anthropic IPO and then drop the nuke 😉 ... if the Chinese plan is to destroy US AI closed labs with open-weight models ...

[Agi Man @AgiMan_WCL]

@AmasCreates No, it is called digital AGI 😉 But we are not there yet ...

[Agi Man @AgiMan_WCL]

@kimmonismus What distillation? RL is the name of the game 😉

[Agi Man @AgiMan_WCL]

@KyleReidhead That's not 'new' money - that's money that will 'go away' from Microsoft, Google, Meta, ... 😉 If they don't pop AI datacenters build-out bubble and Anthropic/OpenAI before ?

[Agi Man @AgiMan_WCL]

@MiaAI_lab By throwing money at Anthropic to boost IPO ($26B?) will certainly not help Microsoft in long run 😉

[Agi Man @AgiMan_WCL]

@GaryMarcus He did something. He pressured Jensen to put some money to his (admin?) coffers for each chip sold to China 😉 But this was already August 2025?

[Agi Man @AgiMan_WCL]

Stop using cloud AI. You are financing your own enterprise demise by this. They will train their models on your know-how and then they will shut you out (like they are doing for AI research) and at the end they will take over your business. Use local models!

[Zephyr @zephyr_z9]

Direct attack at Microsoft

[Agi Man @AgiMan_WCL]

@zephyr_z9 Stop using cloud AI. You are financing your own enterprise demise by this. They will train their models on your know-how and then they will shut you out (like they are doing for AI research) and at the end they will take over your business. Use local models!

[Agi Man @AgiMan_WCL]

@kimmonismus Stop using cloud AI. You are financing your own enterprise demise by this. They will train their models on your know-how and then they will shut you out (like they are doing for AI research) and at the end they will take over your business. Use local models!

[Agi Man @AgiMan_WCL]

@unusual_whales Anthropic will move into new domains (biology, ...) and make sure that their models will be stripped of biology knowledge, to prevent competition in biology domain to catch up. The same, they are already doing for AI research! Use local models!

[Agi Man @AgiMan_WCL]

@unusual_whales Enterprises can use Anthropic cloud AI and will be with this financing Anthropic to train better models for new verticals/domains, and in few years Anthropic will eat the same enterprises. Use local models and finetune them if necessary!

[Agi Man @AgiMan_WCL]

@JohnTinsmanAOT This has nothing to do with 'cancel culture' 😉

[Agi Man @AgiMan_WCL]

@jsnnsa (small and efficient) DNN systems that can 'specialize themself' are the future 😉

[Agi Man @AgiMan_WCL]

@CodeTiwg @mattshumer_ Yep, frontier models will always lead and will always be better at 'everything'. But for certain tasks local models are already good enough. And if one cares about IP rights, privacy, ... then local models are only option. I also use GPT-6 Astra for non-confidential tasks.

[Agi Man @AgiMan_WCL]

@2ALEFTY @mattshumer_ 1.3M+ LoC of C++20 code and Qwen 3.8 Flash Next does good job at code refactoring, planing of new features and redesigns, extracting and documenting patterns, ... Looks like the model was E2E RL-ed (excellent tool use, verifying results, ...) for above use cases to the max ...

[Agi Man @AgiMan_WCL]

@mattshumer_ With 35+ years of experience in architecting and coding enterprise SW systems (multiple verticals/domains) ... I should be able to tell what is good enough 😉 (at least for backends). For e.g. quantum physics research I do agree that local models are useless ... 😉