You’ll Be Surprised: Bard vs. GPT-4 - Which LLM Gets You Better Results?
Can Google's Bard generate results that are demonstrably better in quality and versatility than OpenAI's popular GPT-4? Find out in this Bard vs. GPT-4 battle.
GPT-3 is one of the most versatile and transformative components that you can include in your framework, application or service. However, sensational headlines have obscured its wide range of capabilities since its launch. Let’s take a look at the ways that companies and researchers are achieving real-world results with GPT-3, and examine the untapped potential of this 'celebrity AI'.
The interpretive powers of the GPT-3 autoregressive language model stirred up popular tech headlines when it was unveiled by OpenAI in 2020. It could apparently compose feature articles, write poetry, and even talk to the dead.
GPT-3 has perhaps been a victim of its own publicity; it's a vanguard product in the NLP space, but widespread public interest in the possibility of an effective AGI (which GPT-3 is not), combined with tech media's determination to coopt any new AI product into its annual round of febrile headlines, has left GPT-3 rather misunderstood.
Subsequently, its core capabilities have inspired a slew of startups across a range of sectors.
Here we'll take a look at the usable scope of GPT-3 for business purposes, and at some of the companies that have taken up the vanguard in this respect.
First, let's a look at the strengths and weaknesses of the various methods by which GPT-3 answers a prompt.
Models available for transformations in GPT-3 include Davinci, Curie, Babbage and Ada, each of which have different capabilities in terms of speed, quality of output and suitability for specific tasks.
For instance, Davinci is the most sophisticated of the available models, and is most likely to produce usable output that's more complex, and that explores (or at least appears to explore) higher-level domain thinking and analysis (later we'll also take a look at Davinci Instruct, which is capable of following more specific commands regarding the formatting and domain-specificity of its output).
However, some of the leaner and less computationally demanding models are more than adequate for simpler prompts, saving latency and API request costs. For instance, novelist Andrew Mayne has found that much of the most wide-spread knowledge available to GPT-3 is accessible across lower-level models than Davinci:
Treated as a straightforward 'global oracle', GPT-3 is subject to the same inaccuracies and inexactitudes as the publicly available content that it was trained on, and the depth and truth of its responses on any subject is in proportion to the subject's representation (and misrepresentation) on the internet, and in the standard datasets that informs it.
For instance, Davinci knows a little about Charles Dickens' output, but will give up if you second-guess the first response:
Answering the same question 'Who wrote the novel Bleak House?', the other models seemed less well-read, even on multiple attempts:
But within more narrow domains, usually away from subjective arts and culture topics, GPT-3 proves extraordinarily adept.
The GPT-3 model was developed through unsupervised training, wherein it had to discover, rationalize and categorize vast volumes of data without being explicitly told what any of it 'meant'. In the course of training it was necessary for the model to distil labels from the available entities – such as picture captions and descriptions – that accompanied the content, or that had some kind of discoverable relationship with it.
Where a subject has a rigid and agreed taxonomy and hierarchy (such as mathematics), it's much easier to be certain of the relationships between sub-entities in that domain, and to create generative rules that are accurate.
Programming falls into this category, and GPT-3 is very good at it.
Considering that even professional coders may have eccentric solutions to programming challenges, and that the development community frequently comes into conflict regarding the correct answer to a problem, it's not clear why GPT-3 errs so little when writing code — unless, perhaps, that it might be the only coder who took the time to comprehensively RTFM.
It's also possible that in the lonely course of its training, the model has assigned higher weight to inbound links to 'best answer'-rated posts at sites such as StackOverflow, in much the same way that a search engine's algorithm will calculate authority from frequent citations.
However, GPT-3 is often not the only contributor to those cases where it has won headlines for its programming prowess.
In the case of Sharif Shameen's Debuild, which leverages the OpenAI model to provide instantaneous app development from text input, the user-facing framework incorporates GTP-3, but was also specifically and intensively trained against content from GitHub and StackOverflow.
Here, GPT-3 handles the interpretive and transformational aspect of NLP requests, but will avoid wide-ranging answers in favor of focused, domain-limited responses that are (apparently) guided by the weights of a separately trained model.
There is a well-established market for any product or process that can explain legacy code – live software that may now have multiple dependencies across version-locked VMs, and which in some cases may have roots as far back as the 1970s.
Technical debt of this nature can force companies to retain otherwise unproductive staffers who have become the gatekeepers of arcane but essential systems, or necessitate massive consultation bills if the originating developers should die or depart the company without adequately documenting their sprawling projects.
The Replit Code Oracle uses GPT-3 to explain what complex code will actually do.
The company, which is based in San Francisco and raised $20 million USD in series A funding in February 2021, has even recruited GPT-3 to explain Replit's intent to 'rewrite all the rules of software development' by making natural language the default interface to code generation.
Since Replit has not released details of any co-training that contributes to its code-explanation algorithm, I tested the Replit challenge in the above image in the basic OpenAI GPT-3 playground, and found that the Davinci and Babbage engines can calculate a correct answer by default, without additional domain-specific training:
In developing a complex code interpreter with GPT-3, it is important to remember that where a problem has not been directly assimilated from dataset material at training time, GPT-3 would likely fall back to a text analysis and semantic deconstruction of the written challenge, in the absence of examples in its database that match a particular text pattern.
However, since GPT-3 is very capable of deconstructing query and request logic based on similar taxonomies in a programming language, its 'logical guesses' are well-informed, and likely to be correct.
Research into text-to-SQL predates the advent of GPT-3; but since standard SQL query language is in itself designed to conform to natural language, GPT-3 is very good at reproducing functional SQL requests:
More complex SQL queries are far less human-readable, and nested conditional phrases make mistakes more likely when humans are writing out the commands. Due to the inflexibility of the SQL language, and its restricted taxonomy, GPT-3 has assimilated SQL brilliantly.
New York-based analytics startup SeekWell has adopted GPT-3 as an SQL 'translation' layer.
The company has noted that the Davinci instruct model performs best at SQL queries, since it allows the user to provide specific guidance, such as 'Only respond in correct SQL syntax'.
Currently in closed alpha, OthersideAI's email generation system uses GPT-3 to generate email responses based on bullet-points that the user provides.
Entitled 'Quick Response', the synthetic replies take into account the content and context of other emails in the chain, but seem to transform the responses primarily based on GPT-3's understanding of business email style.
The AI can infer the temporal tense from brief prompts, for instance turning 'good to meet' into 'It was good to meet you'.
However, it’s not clear whether Quick Response could draw on earlier emails that contain meeting details to develop a response such as 'It was good to meet you for coffee Tuesday at Starbucks on 1st St.', or whether it understands how mechanistic and data-stuffed a response like that would seem, and would favor a conversational style instead.
Quick Response is posited as the first stage in a multi-AI communications framework wherein, effectively, two people may exchange information and ideas programmatically and without generating the text content themselves. It's analogous, arguably, to two business entities communicating through their lawyers instead of directly.
At the time of writing, the wait list for the OthersideAI email generation API stands at 15,000+.
Other email-based GPT-3 tools have arisen: Lean To provides a dedicated user interface for AI-based email generation.
When setting up a new project (i.e. an email to single or multiple participants), you can choose the objective of the email, even if it is just 'getting a response', and can import contacts via a CSV file, or add them manually.
As is often the case, the GPT-3-based interface requires a lot of prior information. In this case I chose to write to a Stanford Vision Lab professor, and had to do a fair bit of research in order to fill in all the necessary fields.
The results are unconvincing, and generic enough to likely trigger spam filters, while subsequent attempts do not improve the quality. Despite having put research leads into the contact details (such as a website for the recipient), there's no evidence that any of this information was used.
Bar the odd missing comma, the basic GPT-3 playground does a far better job via the Davinci instruct engine:
The lack of client integration in Lean To is also a hindrance; though the email can be sent through Lean To's own SendGrid system, this would make replies problematic, and increase the chances of the message being automatically flagged as spam. Therefore it's better to just copy and paste the generated output into your own email environment.
Lean To's wizard-driven interface attempts to deconstruct the semantically important parts of a GPT-3 prompt and reassemble it into the 'ideal' prompt (which is not shown to the user), rather than writing it naturally; but 'raw' GPT-3 does a better job of this deconstruction by itself, so long as you choose the right engine and phrase the prompt succinctly.
Ultimately, the value of a GPT-3-derived email generation system is in being able to use it directly in an email client, since, as we've seen, it is easy enough to appropriately prompt GPT-3 in any testbed and ask it to generate an email.
Once GPT-3 has inferred the semantic intent of a text, it can rephrase that content in a number of ways, including ELI5. This is very useful for locating the devil in the details of verbose EULAs and general legal documentation:
GPT-3 has a mode called 'summarize for a 2nd grader' that makes a broadly fair assessment, for instance, of the dense EULA for the videogame Far Cry:
Despite the context of 'high-precision language', legal terminology often contains ambiguous terms that seem designed to produce a 'chilling effect' without specifying what a term signifies, or what a breach of it might entail. Since such clauses are designed to be tested in court, neither a person nor an AI can hope to extract their exact meaning, because, in the context of the document, that has yet to be established.
In the above case, GPT-3 has sided with the cynics, and assumes that paid upgrades will be necessary in order to continue to use later versions of the videogame (rather than any sequels to it). However, GPT-3 has not correctly understood that this probably applies to cases where the existing game is subsequently ported to next-gen platforms, and must be paid for all over again, because the original terminology ('certain hardware') is very vague.
The GPT-3 summary also states that the user can't play the game on any other computer, which is usually not true. Here GPT-3 seems to have mistaken the intent of 'non-transferable' (which means that the user cannot re-sell the license to someone else, and not that the user cannot de-authorize the game on one device and re-authorize it afterwards on another).
In fairness, a lot of actual people would struggle to define the meaning of this term, in this context.
It seems likely that a GPT-3-based framework for decrypting legal jargon would benefit from cross-training against a dedicated dataset of legal terms, where manual labeling could eliminate some of these ambiguities, or at least give GPT-3 enough confidence to declare that certain legal terms may not be clear in their intent, and to outline some possible meanings for them, based on historical examples.
Though the AI-generated content sector was already populated with platforms such as Articoolo, Article Forge and Phrasee, GPT-3's text-generation and semantic capabilities inspired a slew of new AI content startups around the time of its launch, and since.
Some of the more notable names include LongShot, CopySmith, WriteSonic, Headlime and PepperType. Though we don’t have space here to trial all of the new GPT-3 content platforms, we'll shortly look at a couple of the most popular.
In a world where an inappropriate or inconsiderate tweet can end a career or tank a company's stock, there's no realistic prospect that short-form social media output from GPT-3 will ever become entirely programmatic.
However, this more succinct kind of auto-generated social media output is what the current crop of companies are really pushing, perhaps because GPT-3 can produce very apposite results in such a short window of attention.
In the long term, the value of GPT-3 copy generation frameworks may lie partly in the (nascent) potential to write more substantial works, such as coherent blog posts (or at least, in providing better article-spinning transformations than are currently available); in producing corrected or alternative versions of shorter, human-generated social media snippets; and in generating correct English language short-form output for users that have a different native language.
Founded in October 2020, CopySmith leverages GPT-3 to provide a suite of templates to massage existing copy and generate new text-based material from limited prompts.
CopySmith offers text-completion, rewriting and generation services for a range of contexts, including:
…and many more besides. However, in practice nearly all of these possible use cases are minor variations of four basic NLP capabilities in GPT-3:
CopySmith also has a blog post generator, which is currently in beta – but you'll need to 'get out and push' quite a lot:
As with CopySmith's rivals, it's not possible to prompt the blog writer with 'Write me an article about this widget company'; rather, the module focuses on user input, and needs a meaty minimum of 120 words to get started, as well as the initial sentence of the subsequent three paragraphs. However, it is possible to generate these requisites in other CopySmith modules, and then input them at this stage.
I filled in the initial paragraph without giving it any real information about the product, besides the brand being advertised and some general information about the target age-ranges for various products, and tested the blog-creation interface:
The texture of the post is consistently upbeat, as one would expect, but the module struggles for continuity of theme, and invents product features and purposes out of whole cloth. Even within the individual paragraphs, where one would expect reasonable continuity, the text struggles to stay within its own proposed themes, and the net effect is incoherent.
CopySmith is planning to develop SEO principles into its GPT-3 output, as well as paired training on industry-specific topics, which may help some of the 'focus' issues in longer sections of text output.
UK-based Writesonic offers a very similar range of services to CopySmith. The beta article writer, like CopySmith's, requires hand-written prefacing text, between 100-150 words.
Initially, the system requires a target topic:
Subsequently you'll receive ten article ideas:
After this, it's time – inevitably – to provide a detailed corpus of words from which the article will be manufactured:
Having chosen one, you can then choose from ten proposed article outlines, where every two sections cost one credit:
Finally, you'll get a completed article:
Given that the text prompt was incredibly generic, and designed to challenge the AI, Writesonic has produced a surprisingly coherent article from the draft outline, working productively through the header topics to an industry-standard closing paragraph, including a call-to-action.
WriteSonic gives you 10 free generation tokens, but just one attempt at an article is likely to burn through all of them, and so it is very difficult to give the product a second chance if it disappoints.
Though the GPT-3 article writing system AI Writer claims that the content-theft sentinel CopyScape identifies 94.47% of its articles as unique, CopyScape could not find any matches for the final article texts generated from my experiments with CopySmith and WriteSonic.
One word of (anecdotal) warning: when GPT-3 includes interview quotes in its article-style output (i.e., when I have asked for text, but not specifically for quotes or for interview material), I've found, in my two months of experimenting, that not even one of those quotes was available on the internet.
Since GPT-3 is trained on public-facing documents that are likely to be indexed in major search engines, this has led me to believe that although GPT-3 can find apposite interview material, it may currently be reluctant to quote a source verbatim.
I tested this theory in the GPT-3 playground, asking Davinci Instruct 'What did Einstein say about dice?', expecting to receive the famous quote 'God does not play dice with the universe.'
An exact-text search on Google indicates that either Einstein never said any of these things, or that if he did, the internet doesn't know about it. These are, at best, paraphrases — a core function of GPT-3.
Further submissions were unhelpful:
This would be less of a hazard if GPT-3 was consistently wrong about quotes; however, it does get many famous sayings correct:
Watch out as well for persistent session variables in a GPT-3 consultation:
Here GPT-3 is 'on a roll' – in an effort to maintain continuity and return improved results as the session develops, it has trained on my previous prompts, and failed to notice that I am now looking for Einstein quotes, and not (as with the previous prompt) for Winston Churchill quotes.
Shortly after the release of API access, New York-based product designer Jordan Singer developed a GPT-3 plugin capable of passing instructions to the proprietary vector graphics editor Figma.
Called Designer, the plugin communicates with the Figma work space, converting descriptions into vector representations of a described shape. Among its other programming capabilities, GPT-3 has assimilated the postscript language, which represents geometry by systematically describing it in a mathematical space. Designer converts plain text instructions such as 'draw a red circle 500px' into a corresponding postscript command, and then wraps it as a JSON to pass it to Figma.
GPT-3 is capable of generating ideas for company startups, with minimal prompting.
In December of 2020 one outlet published a list of genuine startups, mixed in with fictitious startups suggested by GPT-3 (which had been briefly trained on suitable real-life examples in the course of an API session), challenging readers to distinguish them.
Though the best results were included in the feature, the unedited list of GPT-3 suggestions is available on Google Docs, together with the 'real world' examples that informed them. Many of the ideas are imitative of real proposals, while others lack any particular 'angle'.
The IdeasAI site allows subscribers to receive GPT-3 startup ideas, which are generated daily, in an email. However, once again, the pitches are frequently vague, and some of them are simply random excerpts from related web articles.
Nonetheless, it’s reasonable to assume that dedicated training and more targeted data sampling may hold considerable potential for generating more innovative and off-beat startup concepts across a number of sectors.
The kind of domain understanding that GPT-3 offers makes it possible to perform 'elliptical' searches, where even if none of the words or phrases in a query are found, their intent can be discerned, and relevant results obtained.
Algolia's suite of commercial search tools incorporates a number of machine learning technologies, including GPT-3, to facilitate semantic search on websites, or on any corpus of information that can be indexed.
There's nothing new about semantic search in itself – In a 2013 update to its search algorithm, entitled Hummingbird, Google began to leverage its own semantic Knowledge Graph, which had debuted the previous year, and which operates similarly to GPT-3 in this aspect.
Though the search giant did not entirely abandon keyword/keyphrase frequency as an index of relevance (and has not abandoned them since), it was clear that the future of search might lie in generalized domain understanding, where a lexicon of related terms and images would constitute a 'cloud' of essential information on a topic; and not in merely parsing text and counting occurrence frequencies.
What is new is that accessible and increasingly affordable machine learning infrastructure makes it possible for smaller players to implement truly effective semantic search, either for the purpose of developing new generalized search engines based on web-crawling, or for more specific functions, such as recommender engines.
The above movie and TV recommendation system utilizes GPT-3 and metadata from the IMDB.
The system can also identify a movie from a description that may contain few or no keywords in the target entry:
For entry-level startups, the democratizing power of AI-driven semantic search via GPT-3 is perhaps one of the most exciting aspects of OpenAI's language model.
GPT-3 is a powerful tool that, intelligently applied, can help create genuinely useful and innovative new products and services that were not financially or technologically feasible before.
Though its core strength lies in its power to manipulate domain knowledge (without entirely abstracting it), some of the critical publicity it has received since launch has instead focused on what it does not yet do well, such as generating long and coherent treatises with minimal supervision or input.
However, that's a far broader challenge for the NLP community, and one that's ongoing; nor did GPT-3 ever advertise itself as a novelist or long-form feature writer, though it does have some nascent ability in these fields.
Plugging a raw GPT-3 API call into a new product will work best if your service concentrates on those aspects that GPT-3 can perform well without additional training, such as programmatic analysis. However, there's no real value proposition in discovering and re-packaging GPT-3's native capabilities, since a) it's trivial for any competing product to do likewise, and b) native calls can still be unpredictable, even when querying in a limited domain.
The real potential of GPT-3 is as an adjunct or enabling technology for proprietary systems that train GPT-3 against other dedicated datasets that support your product — as evidenced by the best of some of the services that we've taken a look at in this article.
Can Google's Bard generate results that are demonstrably better in quality and versatility than OpenAI's popular GPT-4? Find out in this Bard vs. GPT-4 battle.
Find out how ReAct prompting enables you to introduce human-like reasoning and action planning into your LLM-assisted business workflows for better results.
How good is OpenLLaMA? Is it something you should consider if you're looking at LLMs to streamline your business workflows? Let’s take a look at how it does on common production tasks.
Explore the world of high-resolution image inpainting with the AOT-GAN model. This blog post delves into the challenges, solutions, and impressive results of the AOT-GAN model, providing a comprehensive guide on how to install, use, and evaluate this powerful tool for image inpainting
Explore how the implementation of Fashion CLIP, a domain-specific AI model, improves product similarity search in e-commerce by understanding complex product attributes, enhancing image-text alignment, and providing more accurate search results.
Explore how DollyV2, a new large language model that's free for any type of commercial use, works and fares on language tasks like chatbots and medical reports.
LLMs can generate more accurate answers by reasoning out difficult problems and questions. Find out how to unlock this power using chain-of-thought prompting.
Explore how the Vicuna and FastChat models from LMSYS fare on various language tasks and learn how to fine-tune and deploy them for your business needs.
From customer reviews to medical reports, find out how GPT-4 zero-shot summarization has the ability to streamline document understanding workflows.
Explore how OpenAI's GPT-4 can be utilized for meeting summarization using zero-shot and aspect-based approaches. Learn about the use of multi-task prompting, abstractive summarization, and modern chunking architectures to distill valuable insights from lengthy meeting transcripts. Discover how aspect-based summarization can cater to various domain requirements, offering unique perspectives from the same transcript.