From: redpointai

The landscape of enterprise AI adoption and usage is rapidly evolving, presenting various models for deployment and unique challenges for organizations looking to integrate Generative AI (GenAI) solutions [00:00:38].

Current Deployment Models

Currently, there are several observed models for GenAI deployments in enterprises:

  • Consulting-led [00:00:51]: Large consulting firms like Accenture come in to build custom solutions [00:00:52].
  • Forward-deployed engineering/product [00:00:55]: Similar to the Palantir-esque model, involving engineers building products on-site [00:00:57].
  • Out-of-the-box product [00:01:02]: Selling a ready-to-use product that enterprises are expected to implement themselves [00:01:02].

In the long term (5-10 years), a middle-ground approach is expected to prevail, acknowledging that enterprises will still require some degree of support to integrate this new, complicated technology effectively into the economy and various applications [00:01:19].

Challenges in Enterprise AI Deployment

AI agents are unique in that they require extensive access to enterprise data and systems to drive automation effectively [00:01:51]. This presents several challenges:

  • Privacy [00:02:24]: Agents need to access emails, chats, calls, CRM, ERP, and HR software, demanding a degree of access rarely seen in other enterprise software, making privacy a significant concern [00:02:08].
  • Custom Integration [00:02:41]: Each company uses a unique “tapestry” or “mosaic” of software, necessitating custom setup to bring all context together and integrate it into the model [00:02:43].
  • Guardrails [00:04:16]: Due to the high stakes of mistakes, particularly with sensitive data like salary or customer information, significant guardrails are necessary [00:04:18]. While self-serve setup for agents would be helpful, a middle ground is expected where parts are automated but not the entire process [00:03:43].

Successful Enterprise Use Cases for GenAI

Companies like Coher observe what GenAI applications currently have product-market fit within enterprises [00:04:44]:

  • Vertical Applications [00:04:55]: In healthcare, GenAI is being used to simplify note-taking and form-filling for doctors by passively listening to doctor-patient interactions and pre-populating information [00:04:57].
  • General Categories (Cross-Vertical) [00:05:21]:
    • Customer Support [00:05:24]: The technology is ready, and there’s a strong need across various verticals like telco, healthcare, and financial services [00:05:26].
    • Research Augmentation [00:05:47]: Augmenting humans with agents that can perform a month’s worth of research in a few hours [00:05:51]. This capability can make knowledge work, such as for wealth managers, significantly more effective by allowing models to read vast amounts of data and provide robust, cited research [00:06:04]. This use case is considered ready for prime time and will be integrated into every enterprise globally [00:24:08].

The Role of Reasoning Models

Reasoning capabilities are crucial for AI models, allowing the technology to allocate different amounts of energy based on problem complexity, unlike non-reasoning models that treat all inputs similarly [00:07:31]. Reasoning has led to a “complete step change” in model improvement, enabling models to accomplish tasks that were previously impossible due to their ability to reflect, understand failures, and find alternative paths [00:16:16].

Evolution of Model Capabilities

Current models lack the ability to learn from experience, akin to humans starting as novices and becoming experts over time [00:08:46]. This capability, including learning from user feedback, is a clear missing property of intelligence that needs to be developed [00:08:58].

The “scale is all you need” hypothesis is breaking, as diminishing returns from capital and compute are evident [00:09:21]. Future progress will require smarter and more creative approaches beyond simply throwing more money at compute [00:09:33].

Custom vs. General Models

While models are becoming capable of self-developing internal experts [00:10:42], custom models remain important for integrating fundamental context about specific businesses or domains that is not available on the web [00:10:55]. This includes data from manufacturing, customer transactions, or detailed personal health records [00:11:30].

General models are extraordinary, and synthetic data can significantly bridge the gap in specialized domains [00:11:56]. It’s unlikely that organizations will operate with tens or hundreds of models; instead, a handful of customized models may suffice [00:12:09].

Data Labeling and Synthetic Data

Human data remains necessary but is often too expensive to gather at scale, especially from experts [00:13:42]. For example, finding 100,000 doctors to teach a model medicine is not viable [00:13:47]. The ability to teach models general conversation skills allows for synthetic data generation, which can then be applied to specific domains using a much smaller pool of human-provided, trustworthy data [00:14:14]. In verifiable domains like code and math, checking results makes filtering synthetic data easier [00:14:49]. For companies like Coher, an overwhelming majority of data generated for new models is synthetic [00:15:08].

Human involvement is indispensable for evaluation (Eval) [00:13:15].

Cohere’s Approach to Enterprise Partnerships and AI Application Deployment

Cohere is pushing into the application layer with its platform “North,” motivated by seeing customers repeatedly building the same applications [00:18:17]. These applications, often built by internal AI teams (not product teams), could take a year and lacked user-friendliness [00:18:29].

North aims to solve this by providing a consumer-grade product experience that is intuitive, low-latency, and equipped with enterprise features [00:19:01]:

  • Customization [00:19:14]: UI customization, rebranding, custom data connections, tool integration, and even plugging in other fine-tuned models like Llama [00:19:17].
  • Accelerated Deployment [00:19:40]: Sets companies forward 12-18 months in their product roadmap, enabling immediate distribution of technology to employees [00:19:43].
  • Strategic Attributes [00:19:55]: Not locked into one hyperscaler ecosystem, deployable anywhere [00:20:00]. Releases weights non-commercially and supports VPC deployments [00:20:07].

This vertically integrated approach, where models and applications are built together, offers more levers to deliver the desired customer experience [00:20:55]. For example, Cohere’s Command generative model is optimized for use cases needed in North, such as interacting with ERP or CRM systems [00:21:04].

Developing and Utilizing AI Models in the Tech Industry

Team Expertise

Building the best AI products often requires deep model knowledge [00:22:34]. Even if teams aren’t training models themselves, they try to approximate that impact at the model layer [00:22:47]. Companies like Cohere aim to provide this deep model intervention for their partners, ensuring needs are met [00:23:01].

Future Use Cases

Beyond current applications, mundane back-office tasks are expected to come online as infrastructure for automation (like North) is put in place [00:24:22]. This includes tasks in finance, legal, and sales [00:25:04]. Sales, for example, involves extensive research (company strategy, internal conversations, individual briefings) that can be dramatically enhanced by AI [00:25:17].

Consolidation vs. Federation of AI Solutions

Initially, enterprises may experience a “scattershot phase” where different teams purchase disparate applications [00:26:35]. However, this will lead to an “insane maintenance burden” due to numerous data source connections across different apps [00:27:03]. This will drive a strong push towards consolidation, with enterprises seeking a single platform plugged into everything to manage all automation objectives [00:27:09].

Economic and Strategic Considerations in AI Model Deployment

Strategic Decisions

  • Location [00:27:52]: Starting up in Toronto, leveraging the strong Canadian AI ecosystem (e.g., Jeff Hinton, Ilia Sutskever’s origins), has provided access to top talent and made Cohere a “darling” of the Canadian tech community [00:27:55].
  • International Partnerships [00:29:40]: Cohere actively partners with companies like Fujitsu in Japan and LG in Korea, investing deeply in non-English languages to ensure their technology works well in those jurisdictions and supports local economies [00:29:42]. An open-source project, “Coher.ai,” collected data from thousands of native speakers across over a hundred languages to improve models for everyone [00:30:22].

Differentiation of Foundation Model Companies

In the future, there will likely be a handful of foundation model companies, each finding their niche [00:32:04]. For example, OpenAI focuses on the consumer front, Anthropic excels at code, and Cohere is focused on enterprise and back-office applications [00:31:32]. New generations of foundation models are also expected for specific domains like biology, chemistry, and material science [00:32:13].

Test-time compute requires significant resources, making inference 3 to 10 times more expensive [00:39:06]. However, compute is expected to become cheaper and more abundant per flop [00:39:23]. The emergence of multiple options for training compute and the ability to combine different chip types for supercomputers are positive trends for AI development [00:39:32].

Future Model Milestones

A significant future milestone for models is the ability to learn from experience and user interaction [00:44:46]. This could involve storing interaction history in a database that models can query for context [00:45:47]. Such models would grow with users, learning preferences and improving over time, fostering a stronger connection with the system [00:46:07].

AI Risks and Societal Impact

Concerns exist regarding:

  • Bad Actors [00:48:43]: Potential access to powerful AI capabilities at the state level [00:48:47]. Ensuring liberal democracies gain an advantage is critical [00:48:52].
  • Job Displacement [00:49:04]: The need for infrastructure to facilitate moving and retraining people into new, more fulfilling careers if particular jobs are impacted [00:49:01].

However, mass unemployment is not anticipated, as humanity has “infinite demand,” and the technology is expected to augment people, allowing them to do more and deliver more [00:47:48]. Concerns about existential risks like “Terminator” scenarios are considered less urgent than near- and mid-term issues that require focus from the public and policymakers [00:49:36].