A Familiar Cycle: The Frustration of Almost Solving the Search Problem
August 16, 2024
This essay is the work of a dumb dinobaby. No smart software required.
Search and retrieval is a difficult problem. The solutions have ranged from scrolls with labels to punched cards and rods to bags of words. Each innovation or advance sparked new ideas. Boolean gave way to natural language. Natural language evolved into semi-smart systems. Now we are in the era of what seems to be smart software. Like the punch card systems, users became aware of the value of consistent, accurate indexing. Today one expects a system to “know” what the user wants. Instead of knowing index terms, one learns to be a prompt engineer.
Search and retrieval is not “solved” using large language models. LLMs are a step forward on a long and difficult path. The potential financial cost of thinking that the methods are a sure-fire money machine is high. Thanks, MSFT Copilot. How was DEFCON?
I read “LLM Progress Is Slowing — What Will It Mean for AI?.” The write up makes clear that some of the excitement of smart software which can makes sense of natural language queries (prompts) has lost some of its shine. This type of insight is one that probably existed when a Babylonian tablet maker groused about not having an easy way to stack up clay tablets for the money guy. Search and retrieval is essential for productive work. A system which makes that process less of a hassle is welcomed. After a period of time one learns that the approach is not quite where the user wants it to be. Researchers and innovators hear the complaint and turn their attention to improving search and retrieval … again.
The write up states:
The leap from GPT-3 to GPT-3.5 was huge, propelling OpenAI into the public consciousness. The jump up to GPT-4 was also impressive, a giant step forward in power and capacity. Then came GPT-4 Turbo, which added some speed, then GPT-4 Vision, which really just unlocked GPT-4’s existing image recognition capabilities. And just a few weeks back, we saw the release of GPT-4o, which offered enhanced multi-modality but relatively little in terms of additional power. Other LLMs, like Claude 3 from Anthropic and Gemini Ultra from Google, have followed a similar trend and now seem to be converging around similar speed and power benchmarks to GPT-4. We aren’t yet in plateau territory — but do seem to be entering into a slowdown. The pattern that is emerging: Less progress in power and range with each generation.
This is an echo of the complaints I heard about Dr. Salton’s SMART search system.
The “fix” according to the write up may be to follow one of these remediation paths:
- More specialization
- New user interfaces
- Open source large language models
- More and better data
- New large language model architectures.
These are ideas bolted to the large language model approach to search and retrieval. I think each has upsides and downsides. These deserve thoughtful discussion. However, the evolution of search-and-retrieval has been an evolutionary process. Those chaos and order thinkers at the Santa Fe Institute suggest that certain “things” self organize and emerge. The idea has relevance to what happens with each “new” approach to search and retrieval.
The cited write up concludes with this statement:
One possible pattern that could emerge for LLMs: That they increasingly compete at the feature and ease-of-use levels. Over time, we could see some level of commoditization set in, similar to what we’ve seen elsewhere in the technology world. Think of, say, databases and cloud service providers. While there are substantial differences between the various options in the market, and some developers will have clear preferences, most would consider them broadly interchangeable. There is no clear and absolute “winner” in terms of which is the most powerful and capable.
I think the idea about competition is mostly correct. However, what my impression of search and retrieval as a technology thread is that progress is being made. I find it encouraging that more users are interacting with systems. Unfortunately search and retrieval is not solved by generating a paragraph a high school student can turn into a history teacher as an original report.
Effective search and retrieval is not just a prompt box. Effective information access remains a blend of extraordinarily trivial activities. For instance, a conversation may suggest a new way to locate relevant information. Reading an article or a longer document may trigger an unanticipated connection between ant colonies and another task-related process. The act of looking at different sources may lead to a fact previously unknown which leads in turn to another knowledge insight. Software alone cannot replicate these mental triggers.
LLMs like stacked clay tablets provide challenges and utility. However, search and retrieval remains a work in progress. LLMs, like semantic ad matching, or using one’s search history as a context clue, are helpful. But opportunities for innovation exist. My view is that the grousing about LLM limitations is little more than a recognition that converting a human concept or information need to an “answer” is a work in progress. The difference is that today billions of dollars have been pumped into smart software in the hope that information retrieval is solved.
Sorry, it is not. Therefore, the stakes of realizing that the golden goose may not lay enough eggs to pay off the cost of the goose itself. Twenty years ago search and retrieval was not a sector consuming billions of dollars in the span of a couple of years. That’s what is making people nervous about LLMs. Watching Delphi or Entopia fail was expensive, but the scale of the financial loss and the emotional cost of LLM failure is a different kettle of fish.
Oh, and those five “fixes” in the bullet points from the write up. None will solve the problem of search and retrieval.
Stephen E Arnold, August 16, 2024
Comments
Got something to say?