9 ms·
Google's Gemini model is delayed
- siva7 3y ago> Google representatives earlier this year told some cloud customers and business partners they would get access to the company’s new conversational AI See, that's the difference between OPENai and old g
- razodactyl 3y agoThe rate at which the Transformer architecture can learn and adapt to new information is astounding - so it makes absolutely no sense to me that they can't manage to throw 200 data annotators at their model and make it best-in-class. They did sink capital into Anthropic recently so perhaps that shuffled things up internally / changed some roadmaps? If that's the case - Anthropic's models are extremely capable and can be easily launched to the top of the leaderboards if they go for a less belligerent training approach.
- riku_iki 3y ago> so it makes absolutely no sense to me that they can't manage to throw 200 data annotators at their model and make it best-in-class. these foundation models are trained on trillions of tokens to achieve quality. 200 data annotators with unknown quality issues may be not enough..
- Zolde 3y agoWhat if Google is not behind OpenAI, but waiting for it to catch up? OpenAI got to iterate, but maybe Google instantly move 37'd conversational AI?
- intellectronica 3y agoCould be, but there's a simpler explanation, and absent evidence to suggest otherwise, the simple explanation is preferable.
- mvdtnz 3y agoWhat is "37'd conversational AI"?
- optimalsolver 3y agoThey're referring to AlphaGo's move 37 in Game 2 against Lee Sedol which blew everyone's minds: https://en.wikipedia.org/wiki/AlphaGo_versus_Lee_Sedol https://en.wikipedia.org/wiki/AlphaGo_versus_Lee_Sedol
- aleph_minus_one 3y ago> OpenAI got to iterate, but maybe Google instantly move 37'd conversational AI? If you write such speculative theories, I want to do, too. :-) What if Google already "37'd" OpenAI, but the result is too dangerous to be released to the public, i.e. it takes a huge amount of work to "defuse"/"cripple" their AI for it to be suitable for "public consumption".
- brookst 3y agoOh I want to play! What if Google has an actual GAI that is so good they’ve secretly turned over all strategic decisions to it, and it is directing this “appear to be behind” strategy for reasons known only to itself?
- maxhille 3y ago"Secrectly"? I find "unknowingly" or "accidental" to be more likely.
- ldjkfkdsjnv 3y agoGoogle doesnt want to disrupt its own business model, search no longer makes sense.
- Retric 3y agoLLM’s only cover a fraction of what search provides. If you want anything real time like news the models simply aren’t updated that quickly. Further, booking a flight etc means finding the correct 3rd party aka search.
- brookst 3y agoSearch is just a tool for LLMs. Ask Bing Chat about tomorrow’s expected Starship launch and it will do a decent job. Likewise with 3rd party extensions; a good LLM will make the best choice based on criteria. In a way the LLM becomes the end user from a search/advertising perspective. The actual end user just expects an answer, not a list of 5 ads and 10 possible places to go look for the answer. So yeah, search as an activity still exists, it’s just much less monetizable because the value prop of LLMs is distilling all that noise to a useful answer.
- Retric 3y ago‘A decent job’ is really subjective here. You don’t get to see the pictures from the article it’s summarizing, which may be ok for something as routine as a Starlink launch, but doesn’t help for a hurricane forecast where you want to see its path on a map etc. I rarely just want an answer when I am doing a search. But when I do I generally care more about accuracy than I am willing to trust a LLM.
- brookst 3y agoThis whole product category is barely a year old and you’re pointing out its (real) limitations? If we’re talking about the viability of the search ads business in the next few months, sure, those are good reasons people will stay with traditional search and therefore be monetizable. But as an industry trend, I think search will be 100% LLM intermediated pretty quickly. Pics of a subject, projections of hurricane routes… I’d be very surprised if both aren’t provided by LLMs within 12 months.
- ot 3y ago> Google co-founder Sergey Brin has been spending “four to five days a week” with the developers. I don't see how that can make the project move faster. I wouldn't be surprised it has the opposite effect actually.
- aleph_minus_one 3y ago> I don't see how that can make the project move faster. I wouldn't be surprised it has the opposite effect actually. Possibly. But you surely already realized that business is a lot about "showing off", in this case showing off that AI is a matter for decision at the top level.
- jgalt212 3y agoI disagree. If the group encounters any institutional blockers, Brin can wave them away instantly. Of course, he can slow things down as well presumably because he's not a SME.
- ot 3y agoThey just need him on speed dial for that, not physically present breathing down their necks.
- FartyMcFarter 3y agoI feel like someone with his experience and background should have the self awareness to help avoid that scenario. However I agree it's still a possibility...
- jakevoytko 3y agoI think he believes he’s the ultimate unblocker. Just find all the points of friction and wave a wand and cut through the bureaucracy, force supporting teams to help, force facilities to fix weird problems, etc. A tiny example I heard: he overheard an employee (who was working on some urgent project he parachuted into) complaining that it was hard working late in this particular office because the closest microkitchen had slim pickings by the end of the day. He stops the conversation and asks a few followup questions. For the rest of the project it was always fully stocked with all the best snacks and drinks, milk and cereal always topped off, fruit always available, etc. I don’t remember the exact details (it’s been a decade since I heard this story) but I remember thinking it went above and beyond what I had available close to my own desk. They thought the dinner spread at the closest cafeteria got better too, but they didn’t know if it was a coincidence or if he made it happen. I’m sure there’s a ton of the “the boss said something and we’re overreacting” effect. But so much of Google has ossified and slowed that I’m sure it’s also helpful to have a walking sudo command to make the org work for you.
- devinprater 3y agoI'm not holding my breath.
- kristofferR 3y agoBard is pretty good already, it's often better than ChatGPT 4 for me.
- lovasoa 3y agoReally? In my experience, it's worse than chatgpt 3.5. I had it answer me things like "this model, which costs 38k€ is cheaper than this other one, which costs 29k€". I've never seen ChatGPT make such basic mistakes.
- nycdatasci 3y agoThis is an uncommon opinion. What are example use cases where Bard is superior?
- m3kw9 3y agoBard feels like the little brother trying to out grow the big brother
- throwaway4aday 3y agoI think they're probably tied up in public relations hell. They are too big and too brand protective to risk launching something that has even the tiniest chance of producing content they could get slammed for. They're likely spending all their time on adding layer upon layer of moderation to prevent that from happening. It would be better to just ship and get ahead of it by doing a demo that contains a weird response and use that as an opportunity to educate people on what they need to be careful of when using the tool. They could spin that into them being honest and having their user's interests at heart. Unfortunately, they also have to worry about what their advertising partners think which is the bigger problem.
- m3kw9 3y agoI think it just suck so they are delaying it
- resource0x 3y agoThey need a pause to diffuse the hype thus lowering unrealistic expectations. I can expect a slew of articles to that effect shortly. Otherwise, people will continue to expect miracles, so whatever Google may release will be a huge disappointment.
- startages 3y agoI agree with this, Google has to account for a lot of things before making anything public compared to other smaller players. Otherwise, huge fines are on the way. I hate that some policy makers don't understand anything about technology, but still try to impose rules that don't make any sense.
- crowcroft 3y agoAssuming this is true the impact of this very defensive brand position is that being defensive and risk averse IS the brand. In the long run your actions speak louder than words. I do think it's a bit ridiculous (although quite possibly true) that they would be wary what advertisers think. The ad platform is full of terrible AI options already.
- whywhywhywhy 3y agoGoogle will be a shell of the former glory by the end of the decade and we'll all by baffled at how they fumbled it from such a position of dominance. Honestly the only parts of their brand I can see surviving it are YouTube, maps and Android. Everything else is completely fair game for disruption and some of it is just laying there on the table for the taking (e.g Image search and actually Search in general), heck YouTube could even be replaced if TikTok made a long form feed. We've watched them be total top of the field world wide in machine learning, things they demoed over a decade ago now were basically magic then they proceeded to do absolutely nothing with the tech and just leave it to rot, now look at them scramble to finally try and implement some of that into reality and their organizational structure just isn't up for the task. Current CEO needs to sit in all ML meetings for a week, figure out who knows how to operate and build it, who's stopping it from leaving the labs, separate the implementation wheat from the bureaucratic chaff, metaphorically lock the wheat in a skunkworks for a month and tell them they have to demo in 3 weeks and ship something when the doors are unlocked at the end of the month.
- NullHypothesist 3y agoI do think the one area that Alphabet is still genuinely innovating is Waymo. With Cruise sidelined, if they can manage to turn it into something that makes money (have ridden in SF, and the tech seems to work great), it could be up there with Youtube & Android, maybe even bigger.
- scarmig 3y agoThe simplest answer is probably the right one: Gemini, or at least the Gemini models that Google believes it can serve to customers in a cost effective way, simply isn't up to par. A release where Gemini manages to be at where OpenAI was three months ago would objectively be an achievement, but it'd be a disappointment and disastrous for investor's outlook for Google. So the delay is to kick the can down the road a bit and hope for some miraculous development that substantially improves quality.