Text classification is the most boring LLM feature, with use-cases on every corner
Midjourney v5

Text classification is the most boring LLM feature, with use-cases on every corner

Something that not many companies did before (other than cool ones): large-scale data labeling is becoming ubiquitous with major advancements in generative LLMs, and dirty cheap also.

“ChatGPT Outperforms Crowd-Workers for Text-Annotation Tasks" [link]

“AnnoLLM: Making Large Language Models to Be Better Crowdsourced Annotators” [link]

As the models become as good or better than human writers, a hypothesis that they could also label better than crowd-workers also seems to become true. And that actually feels like a big thing.

If foundational models do still get trained with human feedback (reportedly, the major source of advancements of GPT-3 → GPT-3.5), it’s more unclear if that role still has relevance for simpler scenarios: labeling products, reviews, feedback, prices, etc.

And one could argue, that with the process being (not now, but tomorrow) as easy as connecting Excel files, almost every company will find 1-2-N cases where it could bring value now.

I’ll share a few examples that come to mind.

Commoditization examples:

  • Sentiment score - any piece of communications the company receives/sends would get one: marketing materials, reviews, etc.
  • Support tickets classification and prioritization - just a few prompts away
  • Any free-form string field anywhere would be labeled now
  • Anything that was categorized, but not classified/labeled - as it was expensive and took a long time to do manually (imagine better categories on any site, catalog, or reference system). Even data-advanced companies had to choose where to invest, scale and choice are much cheaper now and available widely.

Businesses that should see some change quickly:

  • Finance, with its heavy data curation (i.e. BloombergGPT case), but it’s actually a call to adopt similar things.
  • Media / Advertising probably will get tools to produce content with very consistent positioning and classify the existing content much better.

Bit of a future (more obvious one though):

Probably even healthcare will see the wide use of it someday.

  • While it doesn't seem to be a consensus yet that LLMs would be hallucination-free, it seems to me as likely to happen. Again, the scale of controlled training data will be larger and larger.
  • So doctor-generated notes would be transcribed, as well as tests. (I think not even one data-labeling company started with that idea, can’t find a source but here’s the enterprise-focused mention from ominous 2020) Models won’t forget to remind of a custom checklist, highlight risk factors, more likely to connect unobvious facts.

Fuzzy personalization systems:

  • Say we have an ideally comprehensive table about customers of the company - every action transposed, calculated, etc. (not that often you meet such.) But not a very big leap to see a rough idea of it when there's a system that could query by 1 key (user_id) when needed.
  • It'd be able to plan 20 different cases of communications, for 20 different personas and plan them to be sent personally. Maybe not now, but it's hardly a leap anymore, say 20 prompts per customer per month - without no-code drag-n-drops, expensive integrations of LTV software, etc. Smart systems are simpler than sequence-oriented ones now.

So yeah, have a look around if free input things are still used somewhere in your company - there could be some hidden gold.

Last, but not least - I do remember to separate solutions in search of a problem and platform shift.

The former is rather a bad pattern that rarely works, the latter enables you to solve important problems which didn’t seem feasible at all.

Those problems could be additional explainability (if you’d better describe products, catalog, maps, etc.), automation/speed (ticket classification), non-formal aggregation, lots of things really, lots to explore and build.

要查看或添加评论,请登录

Maksim Palevich的更多文章

  • Review? The Network State

    Review? The Network State

    Written by a famous Valley Founder - Balaji Srinivasan, it’s a book / a movement / a community targeted for a…

    1 条评论
  • What you do is who you are for tech-competitiveness

    What you do is who you are for tech-competitiveness

    It’s a gigantic and personally fascinating topic. What should/should not happen/happened to EU tech/startup ecosystem…

  • Tech Narratives in one query. “Small-scale” - 1.6B tokens, 20M comments overview.

    Tech Narratives in one query. “Small-scale” - 1.6B tokens, 20M comments overview.

    Many examples in the original Narrative Economics course were based on N-gram viewer - quite a curious tool, in a way…

  • Elo-based choices and a new look at the generated software

    Elo-based choices and a new look at the generated software

    Long story short, I wanted to see if I’d be able to have (not create) some web service with realisation of basic Elo…

  • Review? Innovator’s Dilemma

    Review? Innovator’s Dilemma

    Even though it’s written in 90s, to me it feels like not a bit of relevance was lost. The core research underneath is…

  • Review? When Genius Failed

    Review? When Genius Failed

    First of all, it’s not exactly ‘The Wolf of Wall Street’, depicting lifestyle parts of the financial drama only…

  • Review? The Dream Machine

    Review? The Dream Machine

    I, for one, never knew exactly that the biography (not quite, but close) of a psychologist - J. C.

  • Review? Where Is My Flying Car?

    Review? Where Is My Flying Car?

    I was having another look at it recently and left very happy I did so - there’s quite something about a story of tech &…

  • Review? The Human Network

    Review? The Human Network

    Oh, it’s a very curious one - book on networks (social) by an economics professor (Matthew O. Jackson) based on various…

  • Review? Measure What Matters

    Review? Measure What Matters

    The book with the subtitle ‘The Simple Idea That Drives 10x Growth’ from one of the first Google investors, praised by…

社区洞察

其他会员也浏览了