The bittersweet lesson.

An advance in AI can make weeks of custom engineering unnecessary. A new model can complete work that previously required a carefully designed workaround.

Our responsibility is to adopt the better approach.

In The Bitter Lesson, Rich Sutton describes a pattern across AI research: general methods that scale with computation have repeatedly overtaken approaches built around human expertise.

His examples include search and learning in chess, Go, speech, and vision. The lesson is broader than making a neural network bigger.

The difficult part is replacing work you invested in when a more general method performs better. An elegant implementation loses its value when a simpler approach outperforms it.

Give general methods room to work

A system built from explicit rules offers control and a clear explanation of its behavior. As requirements expand, maintaining that control can require more rules and more specialist work.

A general learning system learns from data, improves with computation, and applies the same underlying method across different problems.

Modern neural networks put that approach into practice: train a flexible model at scale, allowing it to learn behaviors that would be difficult to specify individually.

Specialist systems remain useful. The strategic mistake is to build around a limitation that more capable general methods will remove.

Language provides a clear example. The GPT-3 research demonstrated one language model performing translation, question answering, and other tasks from text instructions and examples, without task-specific retraining.

One model could support several language tasks through instructions and examples. That reduced the need to develop a separate approach for each type of request.

Biology raises the same underlying question: how much complexity can a model learn if we give it enough useful experimental data?

Deep knowledge of a field’s complexity can make a general solution seem implausible. Testing establishes which patterns a model can learn and where its predictions become useful.

Arc’s virtual cell work uses large datasets of cells and experiments that alter genes or expose cells to treatments. The model learns to predict changes in gene expression.

Perturbation experiments measure the effects of specific changes. They provide information about those effects that observational datasets alone do not establish.

The objective is to learn patterns across biological complexity and use predictions to guide further experiments. A model can support useful discoveries without explaining every part of the system.

Better data, more capable models, and rigorous testing expand the work these systems can support. That is the direction we expect progress to take.

Stay general

We expect general models to keep taking on work that once seemed to need its own specialist system.

A model that writes code, reads documents, and reasons about images provides a common foundation for different jobs. Each implementation can build on those capabilities.

On Robocurve’s stationery tasks, Astra completed seven of 100 trials; the specialist MolmoAct2 completed none. Both struggled, and test conditions differed, but the general model held its own.

In DrivingBench, GPT-6 Astra completed a low-speed cone course in a real car on its second attempt, with a human ready to brake.

Completing one supervised course does not establish driving competence or superiority to specialist systems. It demonstrates why the scope of general models deserves continued testing.

We’re building Rifoa for increasingly general, eventually superhuman AI. Our long-term bet is intelligence that exceeds human performance across the intellectual work businesses depend on.

That includes AI contributing more of the research and engineering used to improve AI itself. We expect that contribution to accelerate the development of new capabilities.

This conviction shapes Rifoa today. We continue expanding the work we attempt and updating our methods as model capabilities advance.

We’re building a general AI services company around general models. Our scope spans finance, operations, sales, and work across departments. We apply that ambition across industries.

Bring us the digital work you want completed. We evaluate the difficult requirements, test current capabilities, and expand the implementation as the models improve.

Make progress useful

A new model release creates value when it changes how your business operates.

Adoption requires understanding the job, connecting the systems, testing exceptions, and establishing whether the new capability improves the result.

The gap between model capability and business adoption is a substantial opportunity. Model performance can advance while the same processes remain manual inside a business.

Our responsibility is to turn those advances into better operations, better services, and experiences customers could not get before.

Your people provide essential context: which supplier needs attention, why an approval exists, and what the completed work must deliver.

We combine your team’s operational knowledge with our applied AI engineering. Together, we remove bottlenecks and build systems around the requirements of the business.

That collaboration takes an idea into daily operation. It requires close work with your team, sound model selection, and continued responsibility after deployment.

We retain the business context and replace components that limit progress. That includes retiring implementations we invested significant time in developing.

We measure success through your company’s capability, even when achieving it requires a simpler implementation.

Established businesses already have customers, relationships, and years of operational knowledge. Helping them become AI-native is an opportunity to improve how a large part of the economy works.

They also have distribution, data, and a position in their markets. Those advantages become more useful when AI improves the products, services, and operations behind them.

We expect some of the biggest winners of the AI era to be companies whose business has nothing to do with selling AI.

An applied AI team works with your specialists to deploy these capabilities. Your business gets useful systems without having to build its own research lab.

Rifoa is an AI-native services company bringing advances in AI into the operations of businesses people depend on every day.

Our ambition is to become a defining company in this field, with lasting client relationships around the world.

We are building toward a global business serving clients across industries and countries. Our focus is completed work and lasting improvements in how those businesses operate.

We want to help build companies that last for generations, with knowledge and operations that survive changes in individual employees, departments, and leadership.

That requires engineers with initiative, strong judgment, and ownership of the result. We want an environment where they learn through implementation and remain responsible for the work they deliver.

Bringing AI into the businesses that support everyday life is important engineering work for the years ahead. We intend to earn a position as a trusted company doing that work.

We apply the same approach inside Rifoa. As models improve, we put them to work in our own operations and use that experience to inform client implementations.

Stay for the long term

An AI implementation is built with your team. The work reveals exceptions, process changes, and new applications. The exact scope is not always apparent before implementation begins.

Trying to specify every future decision in a contract can make the relationship less adaptable. The work needs room to respond to new information.

We establish the starting objective, cost, and responsibilities. As the work develops, we agree the next steps together, with clear decisions about changes in scope.

You’re choosing the people you’ll call when the plan needs changing.

The choice involves judgment and instinct. Look for people who understand your business, communicate honestly, and make sound decisions when the answer is not yet clear.

Choose a team whose working style fits yours. You need a partner who explains developments, identifies implications for your business, and works with you on the decisions ahead.

There is a leap of faith in choosing a long-term partner. We intend to earn that trust through delivery and sustain it for decades.

We work with clients through thick and thin: clear about our capabilities, responsible for resolving problems, and committed to helping the business adapt.

We make the additional effort a difficult problem requires. When an approach fails, we acknowledge it, change course, and remain involved until the agreed work is resolved.

Integrity requires honest advice, including limits. We raise concerns early when a requirement is beyond our capability or the plan needs to change.

Our working standards are speed, ownership, and early action on problems. We give people the independence to act and make responsibility for their decisions clear.

Clear governance, honest reporting, and commitments kept from the beginning create the foundation for a lasting relationship.

Both sides should want the other to win

Long-term work requires give and take. Individual decisions will sometimes benefit one side more than the other. The relationship must create value for both over time.

Evaluate decisions in the context of the whole relationship. Negotiating every change for the greatest immediate advantage undermines the flexibility that shared work requires.

Our decisions must support your company’s success, including when delivery requires more effort than expected. Commitment becomes clear through those decisions.

The same principle applies to employees, suppliers, and partners. Their continued success supports investment, better work, and a stronger business for everyone involved.

Commercial terms need to support both companies. Each side must have room to earn, invest, and continue providing value to the other.

Reputation develops through how you work with people. Fair dealing, honest communication, and investment in a partner’s capability strengthen the relationship for future work.

We want clients to continue choosing us because the work remains valuable. You retain the ability to take over what we build and the freedom to decide who maintains it.

The sweet part of the bitter lesson is that there will be more we can do for you.

We intend to keep finding it.