The release of Bard rules that out in my opinion. No one would unveil a product that's clearly worse than their competitor, damaging their brand, their share price, and the morale of the entire staff, if they had a better version ready to launch soon.
The only way this could be true is if Google has more than one AI division and there's no comms between them. Which, I guess, is possible with Google given their competing chat products.
Every fucking other company in the world doesn't bother with this shit. I was able to use OpenAI no problem. I was able to try Bing, no problem.
Bard? Nope, I have no idea if it's good or bad. I can't trust what I read on Twitter because they write all sorts of nonsense about GPT too. You get a sense about what that thing is only if you try it out.
If Bard had been released to the public before ChatGPT, I probably would have been amazed by it (I know people love to point out errors, hallucinations, stuff AI gets wrong, but there's also the huge "Guys, you're having a fully natural, back and forth conversation with a computer. You can reference previous comments and ask it specifics about those comments. This was strictly the realm os SciFi until relatively recently.)
Compared to ChatGPT, especially the GPT-4 version, Bard is nowhere close. I asked it some programming questions, and ChatGPT can write the code and explain it to me. Bard basically gave me some StackOverflow summaries. Even in areas not related to programming, ChatGPT just feels "smarter", that is that it gives more generally correct, insightful responses, where Bard feels more like it is just scraping stuff it found on the Internet and summarizing it.
Note Bard is much more up-to-date than ChatGPT (I can ask it questions about current events), but even then it feels more "superficial". For example, I asked both (for ChatGPT was on GPT-3.5) "What was the outcome of the most recent US midterm elections?"
Bard answered "The Republicans won control of the House of Representatives, while the Democrats retained control of the Senate. The Republicans flipped 22 seats in the House, giving them a 222-213 majority. The Democrats flipped one seat in the Senate, giving them a 51-49 majority. The 2022 midterm elections were the first midterm elections under the presidency of Joe Biden. The results of the elections were seen as a repudiation of Biden's presidency, and they have given the Republicans a strong position to retake control of the White House in the 2024 presidential election."
I don't mean to get too much into political opinion on HN, but that last sentence just seems wrong to me. Even most Republican operatives were hand-wringing about Republicans underperforming expectations in the 2022 midterms. I asked Bard a follow-up, "Why do you think "The results of the elections were seen as a repudiation of Biden's presidency?" I recall even most Republican operatives thinking that Republicans underperformed expectations, for example there was no "red wave" that was widely expected."
Bard responded with what is essentially gobbledygook: "The results of the 2022 midterm elections were seen as a repudiation of Biden's presidency for a number of reasons. First, the Democrats lost control of the House, which is the chamber of Congress that is most closely tied to the president. Second, the Republicans flipped a number of seats in the Senate, which is the chamber of Congress that is more insulated from the president. Third, the Democrats underperformed expectations in a number of races, which suggests that voters were not happy with the direction of the country.
There are a number of reasons why the Republicans underperformed expectations. First, the Democrats were able to capitalize on a number of issues, such as inflation and the COVID-19 pandemic. Second, the Republicans were not able to unite their party around a single message. Third, the Democrats were able to mobilize their voters, while the Republicans were not.
The results of the 2022 midterm elections are a setback for the Democrats, but they are not a death knell for the party. The Democrats still have control of the Senate, and they will be able to block some of the Republican agenda. The Democrats will also be able to use the results of the elections as a rallying cry for their voters in the 2024 presidential election."
I mean, "the House, which is the chamber of Congress that is most closely tied to the president" - wut? "Second, the Republicans flipped a number of seats in the Senate, which is the chamber of Congress that is more insulated from the president." - but the previous response just told me the Dems "flipped a seat" to increase their margin. "First, the Democrats were able to capitalize on a number of issues, such as inflation and the COVID-19 pandemic." - the Democrats capitalized on inflation???
ChatGPT 3.5 answered "I'm sorry, but my training data only goes up until September 2021 and I do not have access to current events. However, I can tell you that the most recent US midterm elections were held in November 2018, and during that election, the Democrats won control of the House of Representatives, while the Republicans retained control of the Senate. This resulted in a divided Congress, with Democrats controlling one chamber and Republicans controlling the other. It's important to note that midterm elections are held every two years, so there will be another one in November 2022."
You can easily cause chatgpt to radically change its style by prepending the right system prompt. Some people even published many many "jailbreak" prompts that undid the safety priming that openai's system prompt defines.
From what I read it does indeed seem that openai's experience is much more balanced; they generally seemed to have tuned things quite well and the competition is lagging behind.
However it's not clear how much of that is due to fundamental quality of the model or how things have been polished in the last step.
It may sound like an academic nitpick, but it's central to any discussions about whether Google has really been left behind and has lost the grip on AI.
They are not doing a great service to their own bottom line by blocking access to a public beta to so many billions of people
Bard: https://i.imgur.com/m5UGjxU.png
GPT-3.5: https://i.imgur.com/XH1266o.png
GPT-4: https://i.imgur.com/ZcaZ9Qk.png
So basically Bard didn't understand the instructions at all, GPT-3.5 understood that it needs to substitute words but can't figure out the proper context, and GPT-4 not only did almost everything correctly, but it actually went kinda meta and wrote a rationale within the context of the story for the substitutions I asked it to make (which it does on this task more than half of the time, BTW).
Google Brain and Deepmind? Not sure about the "no comm" part, but their HQs are in different countries.
The first time I was at Google (~2010) you could send an e-mail to any engineer in the company and you'd get response offering to collaborate within an hour, and usually a CL under review by day's end. The second time I was at Google (present), it takes multiple quarters to get teams that report up through the same director to agree on who's doing which work.