December brought more choices in AI models, from Google’s multimodal Gemini family to Microsoft’s compact Phi-2 research model. Choice could be useful for businesses, but it also made comparison more difficult. A model announcement, an accessible product and an application that fits a company’s controls are different stages of readiness. The decision is not simply which model sounds strongest. It is which approach can perform a particular job accurately, affordably and within an organization’s operating requirements. These three developments show why that distinction mattered at the close of 2023.
1. Gemini’s introduction widens the range of possible inputs
Google introduced Gemini on December 6 as a family comprising Ultra, Pro and Nano. The announcement emphasized multimodal capabilities and described different model sizes for different environments. Gemini Pro began appearing in Bard, while developer access through Google AI Studio and Vertex AI was announced for December 13. The family announcement did not mean every model or demonstrated capability was immediately available to every business.
The business opportunity was broader than asking another chatbot to write a paragraph. Many workflows combine text with pictures, diagrams or other material: a service request may include a photograph, while a report may include a chart whose labels are essential to interpretation. A sensible evaluation would separate those inputs and ask what the model actually understood. Could it identify a missing label? Would it distinguish a visible defect from a guess? Could a staff member trace the answer back to the supplied evidence?
Business takeaway: maintain a short capability checklist for the specific product and version being tested. Treat a demonstration as a hypothesis to validate using representative business material, not as a blanket promise of production readiness.
Source: https://blog.google/innovation-and-ai/technology/ai/google-gemini-ai/
2. Phi-2 challenges the assumption that every task needs a huge model
Microsoft published Phi-2, a 2.7-billion-parameter model, as a research and development release. Its December research article described results on selected evaluations and discussed the role of carefully constructed training material. This was not the release of a complete, supported business assistant. Benchmark comparisons also did not establish how the model would behave on a company’s documents, terminology or unusual requests.
Still, the research offered a useful planning question: must every task use the largest available system? A narrow classification job, such as categorizing internal requests for review, may have different requirements from interpreting an unfamiliar contract. Businesses considering smaller models should account for setup, hosting, evaluation and maintenance rather than assuming that fewer parameters automatically mean lower total cost. Licensing and deployment conditions also belong in the assessment. A prototype can be informative without being ready to handle customer data.
Business takeaway: describe the task before choosing the model. Evaluate compact models where a tightly bounded job, a clear answer set and human oversight make experimentation useful; do not equate a research result with an operational guarantee.
Source: https://www.microsoft.com/en-us/research/blog/phi-2-the-surprising-power-of-small-language-models/
3. API changes make integration discipline more important
Anthropic announced a beta Messages API in December, along with changes to commercial terms scheduled to take effect in January 2024. The API announcement addressed how applications structure interactions with the model. The terms announcement described expanded copyright indemnity subject to conditions. Neither an interface update nor a contract provision removed a business’s responsibility to understand what it was building and how it used third-party material.
For an organization integrating AI into software, this is a reminder to manage the connection as a maintained dependency. Record the model version, the request structure, the expected output and how failures are handled. Store representative test cases so an API or prompt change can be checked against previous behavior. Someone should also own the vendor review, including applicable terms, data handling and the circumstances in which a human must approve an output. Those responsibilities are easy to overlook when the first prototype takes only a few lines of code.
Business takeaway: a successful AI integration needs a maintenance owner and a change test. Budget for those continuing responsibilities alongside the initial build; obtain appropriate advice on contractual terms where needed.
Source: https://www.anthropic.com/news/expanded-legal-protections-api-improvements
A useful year-end decision
Create a one-page comparison for one valuable workflow: what inputs it needs, what constitutes a correct answer, what information it may access and what happens when it fails. Compare available products against that document. This keeps procurement connected to operating needs instead of the breadth of a launch presentation.
Present-day invitation: To identify an AI project that fits your business and its budget, book an appointment at https://loyti.com/book. Brice Prather, President, CEO & Founder of Life On Your Terms, Inc. (LOYTI), can discuss AI consulting, custom software and automation opportunities around a concrete process you want to improve.