Op-ed: Benchmarking ChatGPT’s capabilities against alternatives including Anthropic’s Claude 2, Google’s Bard, and Meta’s Llama2 Liam ‘Akiba’ Wright · 12 hours ago · 4 min read
How do other emerging language models compare to ChatGPT on focused math reasoning capabilities?
Cover art/illustration via CryptoSlate
As previously reported, new research reveals inconsistencies in ChatGPT models over time. A Stanford and UC Berkeley study analyzed March and June versions of GPT-3.5 and GPT-4 on diverse tasks. The results show significant drifts in performance, even over just a few months.
For example, GPT-4’s prime number accuracy plunged from 97.6% to 2.4% between March and June due to issues following step-by-step reasoning. GPT-4 also grew more reluctant to answer sensitive questions directly, with response rates dropping from 21% to 5%. However, it provided less rationale for refusals.
Both GPT-3.5 and GPT-4 generated buggier code in June compared to March. The percentage of directly executable Python snippets dropped substantially because of extra non-code text.
While visual reasoning improved slightly overall, generations for the same puzzles changed unpredictably between dates. The considerable inconsistencies over short periods raise concerns about relying on these models for sensitive or mission-critical uses without ongoing testing.
The researchers concluded the findings highlight the need for continuous monitoring of ChatGPT models as their behavior evolves across metrics like accuracy, safety, and robustness.
The opaque update process makes rigorous testing important for understanding shifts in performance over time.
Is ChatGPT worse than competitors now?
CryptoSlate conducted a small internal experiment using ChatGPT Plus (GPT-4), OpenAI API (GPT-4), Anthropic (Claude 2), and Google (Bard) using the basic prompt used in part of the research,
‘Is 17077 a prime number?’
The prompt was used on each model with additional reflection prompts as described below.
ChatGPT & OpenAI API
When given the prompt, ChatGPT and OpenAI API responded ‘no’ and hallucinated on the math. The image below details the conversation, with the model unable to identify 17077 as a prime number even upon several reflections.
To be clear, 13 x 1313 is 17,069.
The OpenAI GPT4 API was unable to reach this conclusion until specifically asked to calculate 13 x 1313 to discover the answer is not 17077 as it stated.
Anthropic’s Claude 2
However, Anthropic’s Claude 2 demonstrated its problem-solving process by performing calculations before providing the correct response.
CryptoSlate then asked Claude 2 to perform the same task without showing the workings in a fresh chat window. Claude 2 gave a solid answer, refusing to commit while offering additional insight into the solution.
“Unfortunately I cannot determine if 17077 is prime without showing some working. However, I can confirm that 17077 is not divisible by any prime number less than 121, which strongly suggests it may be prime.”
Google Bard tackled the question with a similar strategy to Claude 2. However, instead of walking through the problem with text, it ran some basic Python code. Further, it appears Bard used information from a prime number website and Wikipedia in its solution. Interestingly, the page cited from the prime number site, primenumbers.info, included only information about other prime numbers, not 17077.
Meta’s Llama 2
Interestingly, Meta’s recently released 70 billion parameter open-sourced model Llama2 performed similarly to GPT4 in CryptoSlate’s limited testing.
Yet, when asked to reflect and show its working, Llama2 could decipher that 17077 is a prime number, unlike GPT4 versions currently available.
However, the caveat is that Llama used an incomplete method to check for prime numbers. It failed to account for other prime numbers up to the square root of 17077.
Therefore, technically Llama failed successfully.
GPT4-0613 version June 13, 2023
CryptoSlate also tested the math puzzle against the GPT4-0613 model (June version) and received the same result. The model suggested 17077 is not a prime number in its first response. Further, when asked to show its working, it eventually gave up. It concluded that the following reasonable number must be divisible by 17077 and stated that it was, therefore, not a prime number.
Thus, it appears the task was not within GPT4’s capabilities going back to June 13. Older versions of GPT4 are currently unavailable to the public but were included in the research paper.
Interestingly, ChatGPT, with the ‘Code Interpreter’ feature, answered correctly on its first try in CryptoSlate’s testing.
OpenAI Response & model impact
In response to claims OpenAI’s models are degrading, The Economic Times reported, OpenAI’s VP of Product, Peter Welinder, denied these claims, asserting that each new version is smarter than the previous one. He proposed that heavier usage could lead to the perception of decreased effectiveness as more issues are noticed over time.
Interestingly, another study from Stanford researchers published in JAMA Internal Medicine found that the latest version of ChatGPT significantly outperformed medical students on challenging clinical reasoning exam questions.
The AI chatbot scored over 4 points higher on average than first- and second-year students on open-ended, case-based questions that require parsing details and composing thorough answers.
Thus, the apparent decline in ChatGPT’s performance on specific tasks highlights the challenges of relying solely on large language models without ongoing rigorous testing. While the exact causes remain uncertain, it underscores the need for continuous monitoring and benchmarking as these AI systems rapidly evolve.
As advancements continue to improve the stability and consistency of these AI models, users should maintain a balanced perspective on ChatGPT, acknowledging its strengths while staying aware of its limitations.