One Model Won’t Rule Them All?
- There’s been a slight vibe shift, lukewarm GPT5 reception, MIT report showing majority of enterprise gen AI initiatives failing, Sam Altman cautioning over exuberance (or looking to talk capital out of the market), scaling computes overblown so now we’re all reasoning maxi’s etc.
- I’m not smart enough to comment, but what does seem apparent, without assuming any additional model improvements we have a 5-10 year cycle ahead to disseminate AI automation throughout the economy. Blessed are the consultants.
- This has been obvious to many, triggering the rise of the AI automation agency to own this last mile. We’re seeing countless young, technical founders building their own agencies to capture the moment. Harnessing this wave has propelled the GTM motion of platforms like Clay, N8N and Vapi. With a new crop of companies emerging, perhaps opportunities arise akin to the PSA systems (Connectwise, Halo, Ninja etc.) or verticalised payment solutions (Flexpoint) which support the MSP ecosystem in IT. We’re already seeing founders starting to build for this emerging ecosystem …
- So as immediate hopes of ‘AGI’ or model generalisation ad infinitum subdue are we in for a more ‘gentle singularity’ and perhaps a medium term of domain specific agents? It seems there is value still in the grunt work of mapping business ontologies to define workflows and data schemas underpinning industries. The last mile of AI looks more like the 100 page prompt KPMG tax agent built using tax advice written by partners stored largely in their heads. So specialised that only tax agents can use the tool. Lucky them!
- And without one model to rule them all Nvidia thinks we should be focusing on small language models. No wonder Apple doubled RAM on Mac’s last year without increasing prices. Long HBM? As optimal agentic designs continue to shake out, there are clear latency, cost and privacy rationale behind leveraging SLMs for routine, narrow tasks and reserving LLMs for more complex reasoning. If the primary task is tool calling the computational weight of an 140 IQ model makes no sense.
- Pausing on Section 6 of Nvidia’s paper regarding LLM-to-SLM conversion, perhaps the enterprise complexity start-ups can help sherpa reveals? Secure usage data collection, curation, task clustering, SLM selection, specialised fine-tuning etc. Should I be using CoT prompting, majority voting or tree-of-thoughts? Wait can’t Sakana & Writer also give us continual learning? Who’s going to hold KPMG’s hand?
Are Grads Cooked?
- Stanford researchers say employment for young people with AI exposed jobs is down 13%, but where AI was ‘augmentative’ are actually increasing
- That definition is based on the Anthropic Economic Index, a classification of 4m claude interactions to determine if queries were augmentative of human capabilities (e.g., learning or iterating on an output) or resulted in automation (e.g., fulfilling a request with minimal human involvement).
- But that distinction seems partly based on how you approach using the tool. Asking Claude to one shot this entire article by feeding in a paper and style guide would be automation, but sparring on its content iteratively, augmentative? Regardless, perhaps that’s the point, right now it’s all a skill issue. Knowing the right questions to ask, having agency and time to explore emerging capabilities to create your own leverage
- Ultimately nobody has more than 3 years applied gen AI experience right now, so being a teenage who can tinker 16 hours a day to push up against the edge of the possible is a super power. Forget sentiment that 18 years olds are cooked, your lack of priors, flexibility to work hard and lower cost base gives you an unfair advantage to learn & drive adoption. If AI is augmentative you’ll leapfrog the luddites, if it automates us all then you’ll be on UBI running ultra-marathons to find meaning anyway. But don’t listen to me:
If you are AI-native right now, coming out of college, the amount you can teach a company is unbelievable – Aaron Levie
Companies don’t understand how to implement AI right now to get a competitive advantage… learn to customize a model, walk into a company, show the benefits. That is every single job that’s going to be available for kids coming out of school. – Mark Cuban
If you’re tinkering around the edges of the possible I’d love to chat – jamie@triplepoint.vc