Frameworks like LangChain decouple application logic to ease swapping
@OpenRouter Big Crunch example of the day:
Pulled the trigger today and switched 100% of Lindy traffic to DeepSeek v4, churning from Anthropic models. Saves us millions of $ and we're actually seeing an *increase* in performance on many core use cases. Transformative for the business.
Positive users praise the major cost savings and performance gains from companies switching to DeepSeek from Anthropic, while negative users question Anthropic's motives around safety messaging and IPO timing.
transitions like this are why we think it's helpful to have a provider-agnostic harness
we used to talk more about swapping models when the latest and greatest came out -- but the latest and greatest from major providers are expensive!
we're seeing more folks swap to open source models for cost reasons, which is easy with deepagents and/or langchain
Pulled the trigger today and switched 100% of Lindy traffic to DeepSeek v4, churning from Anthropic models. Saves us millions of $ and we're actually seeing an *increase* in performance on many core use cases. Transformative for the business.
If you want to understand why Anthropic is crying wolf about the dangers of uncontrolled AI development, this tweet is the best explanation.
Pulled the trigger today and switched 100% of Lindy traffic to DeepSeek v4, churning from Anthropic models. Saves us millions of $ and we're actually seeing an *increase* in performance on many core use cases. Transformative for the business.
@lugaricano No, the founders have been very outspoken about their concerns on ai safety before they even founded the company The whole current AGI boom was birthed by the rationalist "doomer" community "This invention might literally kill everyone" isn't typically the best PR strategy
@lugaricano I'm a bit confused, how would slowed down ai development help them to compete with deepseek? I assume a "ai research slowdown" would be measured in GPU MwH, and the deepseek team is actually great at getting more juice out of limited compute (eg DSA)
@MKelner great question
you definitely have to tweak the harness based on the model ala https://www.langchain.com/blog/tuning-deep-agents-different-models
that being said we make it easy to customize the harness and swap in any model you want! harnesses like claude agent sdk or openai agents aren't provider agnostic :/
@sydneyrunkle The question is does a provider-agnostic harness even exist? I feel like the model have enough different characteristics that swapping out a model without changing the harness in any way will not lead to optimal performance
@Altimor its very good
Pulled the trigger today and switched 100% of Lindy traffic to DeepSeek v4, churning from Anthropic models. Saves us millions of $ and we're actually seeing an *increase* in performance on many core use cases. Transformative for the business.
@lugaricano I don’t know Luis. Seems like the folks at Anthropic have been very consistent about their focus on risk management. @bcherny recently said in a podcast that the one reason everyone shares for working there is their focus on safety.
@lugaricano Since when are anecdotes more noteworthy than revenue numbers? And didn't they care about AI safety independent of their market position, whether as upstart or leader?
@henrytdowling I don't think AI research slowdown is the point. The point is the control. Slowdown is the way to say control without saying that.
@lugaricano @elvillegaschile
@lugaricano You’ve had a couple of pretty bad and unnuanced takes recently. Disappointing.
@lugaricano I agree with @Altimor here, the consistent messaging from anthropic about this, even at times where their material incentive would have been otherwise indicates to me this is not simply strategic posturing for market entrenchment position the interests might align now, for sure
@Altimor @lugaricano The timing to call for a global pause is very clever though. Now that the IPO is underway and they are positioned as industry leaders, they stand to benefit the most from a slowdown
@sydneyrunkle Is it about a provider agnostic harness or a harness agnostic layer?
If do the latter, you can also leverage the 'top harness for a given model'.
If you add activity type to the mix, you'll end up with activity <> harness <> model configurations. That's the benchmark missing.
@lugaricano this is a lazy accusation that says more about your character than theirs
Yup! You guys have been releasing great stuff around this. At the same time I still feel like migrating from one model to another is still a dark art that not only has to take in model characteristics, tools etc. but also how all those things interact with the thing you want the model to do
