Helping organisations leverage AI. Head of Translational Research, Deakin Applied AI Initiative. Robustness, Complexity, Evolution, Knowledge Systems, Advaita.
Melbourne
Joined January 2008
- Tweets1.1K
- Following1.5K
- Followers492
- Likes16.4K
"The deepest risk of the current trajectory is not that AI will fail to replicate expert knowledge. It is that AI will succeed just enough to disrupt the conditions under which expert knowledge forms, while failing to provide an adequate substitute." open.substack.com/pub/rvasa/…
Mind shift change to make given recent advances. AI agent swarms are parallel, concurrent & not sequential; coherent orchestration (taste), judgement (what to ignore vs attend to); constraints & bounds as first-class citizens of design. AI memory may be solved, humans still limited.
The like the approach a lot @tomkrcha. Blending the fluid stuff (linguistic / language interface) with a rigid structures supports evolution & adaptation. Maps closer to how software is grown in practice ... incrementally.
Starting to feel a sense of deja-vu scrolling through new all the news about AI? Does it all seem to be the same stuff over and over about AI across LinkedIn, X, Discord, Reddit? Ever wondered what is going on?
My short answer ~ "the world is full of isomorphisms (coercive, mimetic, normative) & these forces shape the technology landscape".
So, what the heck am I talking about? Longer answer below -- but feel free to take my short answer, paste it into any LLM and ask it to contextualise it for the field if "AI" ~ it most likely should give you what follows; skim away my take to save time :)
In this context, it is the process by which organisations or systems adopt similar structures, practices, or behaviours over time, often due to external pressures rather than internal needs or efficiency. For example, most universities end up with courses, subjects, lectures, labs, exams, assignments, internships, discipline areas, focus groups ~ the structure is similar across universities; even if some give them slightly different names. It’s a way of explaining why institutions in similar environments end up looking and acting alike.
There are forces that pressure organisations (or groups of people) to converge on what will feel like a single, accepted form. These show up a lot in tech with interesting effects across the eco-system (the isomorphism explanation was put forward by DiMaggio & Powell initially).
In AI, the same forces are at play across 3 broad aspects.
Coercive Isomorphism (Constraint from above)
A top-down pressure from external mandates. Consider GDPR, AI ethics principles, or CCPA forcing specific protocols. These manifests as UI design choice and even AI model architectures. Compliance becomes a non-negotiable design constraint before a single line of code is written. The challenge is not longer about novelty, but to innovate within these pre-defined, and often rigid, guardrails.
Mimetic Isomorphism (Pull of leader)
In a field defined by high uncertainty or ridiculous velocity like we see in AI, the default strategy is often to copy the market leader. When a firm like OpenAI or Google put out a new recipe or tactic, the industry reflexively mimics it. Model Context Protocol (MCP) with all its security gaps still takes off, agent protocols from Google are given a significant amount of attention ~ tech media covers these further amplifying the trend. This feels like a safer, more rational bet, but it is an innovation trap & longer term we get technofeudalism (even though no one wants it or voted for it). While the ecosystem is busy iterating on the leaders position, breakthroughs demand divergence, not just incremental refinement.
Normative Isomorphism (Pressure from within)
This force stems from shared professional norms, shaped by top universities/think-tanks, influential arXiv papers, and conference groupthink. We internalise a "standard toolkit" as the "correct" way to engineer even if (sometimes) it makes no-sense in terms of the trade-offs being made. This shared baseline is powerful, yet drives us all into a cognitive rut.
My feel is that the best ideas in AI are likely to come from merging or re-mixing different disciplines ~ yet, this approach is not given sufficient power & hence the groupthink over rides sense.
As engineers, we are the architects of the future, but isomorphisms are like an invisible resistance & shaping force. Just knowing about it helps. So feel free to question constraints without breaking them, resist the urge to imitate, and we all should aim to challenge the orthodoxy of our training (within common sense norms).
Now more than ever we need to make progress in an interesting way and this just will not come from convergence, but from deliberate, informed divergence.
My definition of creativity is the distance from our past & the history of the field.
This is an experiment in attempting to write longer form on X based on ideas that I have encountered over time.
Deep & insightful point @karlmehta re: structural problem with AI in realworld. Without the right structures and systems that actively remove bottlenecks / roadblocks .. we get cute prototypes.
A useful tool & dev friendly secure architecture. Have a peek at the blog. Look forward yo seeing the adventure @_lbowers
LLMs are reshaping software engineering, automating code but not replacing human creativity & problem-solving. rvasa.substack.com/p/softwar…
This is a subtle and deep point; esp. as machines construct queries // we may very well need different interfaces for humans and machines. Also, the decision & action path+velocity of machines is vastly different to humans - this also poses different challenges. UX for for humans has a narrow focus; machines can and will do very different tasks as well.
" the people who win distribution will own demand. the rest will rent it" .. - the gold nugget.
the new startup playbook looks NOTHING like the old one:
– most of your team will be part-time contractors, creators, and ai agents
– your first $1m will come from niching down. your next $10m will come from tastefully scaling out
– one agent spins out 50 longtail SEO pages from transcripts, support tickets, or user reviews
– startups are turning into QVC. except this time, you own the channel and the product
– onboarding will feel like texting a friend. static forms are dead
– every landing page rewrites itself based on who's viewing it (claude or chatgpt-4o + session data)
– every successful company will feel like a subculture. the product is just a portal in
– outbound are agents scraping, qualifying, and writing personalized intros 24/7
– customer support = 1 human backed by 5 lindy agents trained on every support ticket ever written
– micro-apps will outperform mega-tools. specific > general
– growth isn’t an afterthought. it’s built into the product (agent-invite loops, ai-powered referrals)
– if your product doesn't spark curiosity in 2 seconds, it’s invisible
– the best products of the next decade will be memes first, software second
– “launch” is outdated. leak it instead
– the new pricing model: $0 to play, $x to unlock identity
– you won’t sell software. you’ll sell outcomes, transformations, identity upgrades
– more people will leave big tech to build solo. not out of rebellion, but because their side hustles are more interesting
– the best homepages become a scene. your standard shadcn websites won’t hit the same
– default alive is low burn, small team, owned audience, high-leverage systems
– competitor research happens automatically. agents scrape, cluster, and surface positioning gaps
– your CRM isn’t stale. agents log calls, summarize deals, and write follow-ups before you hang up
– venture capital is optional
– customer success isn’t reactive. agents predict churn based on tone in support chats and usage
– we’ll see more “tiny empires”: one founder, one audience, and a constellation of tools they own
– bug reports are summarized, tagged, prioritized, and triaged by an agent before eng ever sees them
– IRL matters. founders become event planners
– most SaaS is overbuilt. the next wave wins by subtracting
– if your product can't be explained in a screenshot, it won't spread
– the creative director is the new power hire. taste is now a growth lever
– churned users get a custom winback campaign built by an agent based on why they left
– knowledge base builds itself from slack threads, loom links, and discord q&a (agents + gpt vision)
– product feedback loops are instant. users speak → agents summarize, prioritize, and mock ui changes
– most startups will die trying to be “all-in-one.” the winners do one weird thing stupidly well
– startup advice used to be: find a technical cofounder. now it’s: find a distribution edge
– your product isn’t finished when it works. it’s finished when people want to wear the hoodie
– the people who win distribution will own demand. the rest will rent it
if this felt like a glimpse into the future, it's because it is.
instead of bookmarking this, share it with a friend, and start building.
you don’t need permission to build like this.
you just need to start.
most people will ignore this.
but this is the new reality...
small teams, infinite leverage.
Happy building.
I'm rooting for you.
Index to actions rather than outcome. This also means, you have to enjoy and be motivated for the action part always. Very interesting take.
Our RAG pipeline probing study also showed the weak points. Good to see this work and progress. We will soon present updated paper at CAIN 2025. Last years version is at arxiv.org/abs/2409.19019
Existing IR/RAG benchmarks are unrealistic: they’re often derived from easily retrievable topics, rather than grounded in solving real user problems.
🧵Introducing 𝐅𝐫𝐞𝐬𝐡𝐒𝐭𝐚𝐜𝐤, a challenging RAG benchmark on niche, recent topics.
Work done during intern @databricks 🧱
The hardware advances alone will help a ton (scaling law assumption). A few architectural advances .. even if incremental will be a step change. The bottleneck will be human ability to fully leverage (at times .. even o3 and Gemini models feel this way).
So close @OpenAI o3 model. More impressive is the reasoning for why the model thinks it made a mistake. Open thread for Prompt and its internal reasoning trace.
