You might be surprised how good cloud gaming has gotten. I play AAA games at max settings on my MacBook Pro through GeForce Now, and with fiber internet it's nearly indistinguishable from native.
I know people in general hated it, but I found Stadia to be quite good. I'm not too upset because Google paid me back full purchase price, but it's almost a shame that they managed to mess up cloud gaming that badly.
I don't know, I saw quite a few positive comments on Stadia, both as a service and the general approach. Most of the negativity was about it being a Google product and not wanting to get invested in a platform they would inevitably kill. Then of course there was the reaction when it was inevitably killed.
I use boosteroid, which is just steam on cloud. ~4k @ 120Hz for $12/month. No HDR though (they recently removed it). Such a stupid good deal compared to the price of a gaming PC, that I can't really complain. So many data centers with GPU sitting around...
Yes, it's just streaming a video to you. The main limit is your connection speed if you're not near a datacenter as you're limited by ping, so controls can be laggy. You can try it out for free though, which will give you an idea of how good your link is.
It's incredible how far behind Gemini has gotten, both the product and the model. Even the ChatGPT plugin for Google Sheets blows away the native Gemini integration.
Everyone thought Google was pulling ahead with Gemini 3. For a minute there they had the best language model, image model, AND video model in the world. But it's like they decided to pull over for a nap while OpenAI and Anthropic flew by.
Maybe they've decided they don't want to play the same game as OpenAI and Anthropic? They're much better positioned for the high volume AI work that's likely to be where the money is made, with calls to APIs doing routine things for all the businesses of the world. They're also the only big US player that has an open model that you can build on. I don't think vibe coding or the most cutting edge capabilities are what will determine profit from AI.
> They're much better positioned for the high volume AI work that's likely to be where the money is made, with calls to APIs doing routine things for all the businesses of the world
How, exactly, are they currently conquering the enterprise world with their models? What do you think Anthropic is doing?
Their latest proper model is a year old, they have no moat, no enterprise commitment.
Your comment would make sense if they would have actual success in the enterprise market and would have actual products in that area, but they don’t.
They had a brief sprint, caught up, and then dropped the ball again.
Their only current moat is their TPUs, and the fact that
1. The whole (successful) LLM world is screaming for capacity
2. They have excess capacity to rent out, just like Grok
What's a "proper model?" Gemini 3.1 Pro was released 3 months ago. Gemini Robotics 1.6 was released a month ago. And Google is vertically integrated, they aren't just selling tokens, they are selling Taxi rides with Waymo. AI is a lot more than LLMs and Google is doing a lot more than LLMs.
> How, exactly, are they currently conquering the enterprise world with their models?
I didn't say they were conquering the enterprise world. I said they are better positioned for the work that will be profitable in the future. Winning will mean being "good enough" for things like routine interactions with customers at the lowest cost to the business, and having customers fine tune your models using your hardware.
> What do you think Anthropic is doing?
Aside from being arrogant jerks that don't care about pissing off their customers, they're positioning themselves as the highest price provider for the highest end work. There will be a market for that, and maybe Anthropic will survive, but Google looks to me like they have a shot at being the profitable AI company.
If you're building on top of APIs and can do some eval work (aka do not need the most bleeding edge model), the Gemini Flash and Flash Lite models are super capable for the price.
I have the opposite experience where Gemini (even the flash models) has the only useful model for my reverse engineering related use case. My hunch is Google utilizes its free access to entire Google search indices to train itself from niche non-English speaking community websites, much frequently and in a "relevant" manner, which in the end gives these models the most up to date info for this particular kind of work. Every other model is just either 10 years outdated with their answers or simply hallucinates like waaaay crazy.
> for my reverse engineering related use case [...] Every other model is just either 10 years outdated with their answers
I've mostly been doing reverse engineering with Codex, mostly related to games, but not once has the "training data cut-off date" been in the way, the most useful part comes from handing it a binary/directory and letting it prod it until it finds the answer you're looking for, I don't even have web search enabled and sometimes it might take 30-40 minutes for it to find the answer, but I never saw it be unable to find the answer because it's training data was a couple of years old.
3.1-pro is still very capable, and API is at competitive price vs e.g. Anthropic, they just can't seem to figure out RLHF and harness. It needs a lot of guiding, it tends to be lazy and poorly sticking to instructions by default.
It just feels like many google products really, they are capable of really amazing things, it's just that nobody there seem to care. I would guess they are likely optimizing more for internal use than their vast userbase.
I just cancelled my Gemini subscription yesterday. I have a big private fork of OpenCode, and I did it the wrong way to start with, so I couldn't pull from upstream.
So I put together a plan for refactoring it, step by step, with tests, etc. After literally 8 solid days of fighting with Gemini 3 Pro, I still couldn't pull it off.
I gave GPT 5.5 a chance with the same prompt, plans, and repo. I'm not sure how long it took, but when I checked in on it a few hours later it was done. All tests passed, everything exactly how I'd asked, and better (it made some improvements).
My non technical wife knows both ChatGPT and Anthropic (admittedly, because of me) but doesn’t know Gemini. This is amazing to me.
Surely she has seen Gemini in Google search but even her use of that is plummeting.
Google has so much revenue that they’ll be around for a long time. But I feel they are fumbling the opportunity with AI. Even in corporate, where we have Gemini. The conversation is fully around Claude. No one talks about Gemini.
Reports of the death of Google Search have been greatly exaggerated.
If you believe all the reports on HN about everyone's non-technical wives and grandmas, you'd have a hard time explaining the all-time highs in global usage and revenue from Google Search.
I agree with you that Claude 4.7 Opus is better than Gemini 3.1 Pro, but it's also a lot more expensive.
For my applications, I can't find better price-performance than Gemini 3.0 Flash. And it hasn't even been upgraded to 3.1 yet.
I suspect Google's target is price-performance and not just raw performance, which is how they can serve LLM responses at Google Search scale and still set an all-time record for quarterly earnings of any public company ever.
Frontier model capabilities leapfrog each other every few months, and Google I/O is in ten days, so I expect the leaderboard will change again soon.
Unfortunately, I think Google is in the process of killing the golden goose. I visit so few unrecognized websites now and primarily rely on “AI mode” to answer my specific question rather than sift through a handful of possibly accurate pages. How long can that go on before those sites just no longer exist and the source of that knowledge or new knowledge evaporates. Doesn’t seem like that model is sustainable long term.
Honestly, I think the SEO virus killed that golden goose long before the first AI chat bot. If we still had good search taking us to sane websites, ChatGPT might well have never been a thing. I was posting (including on HN) about the vulnerability of Google's search business years before AI chat. It just happens to be the thing that filled the gap when usable search disappeared.
OpenAI and Anthropic have no moat. DeepSeek is a drop-in replacement that is really close in performance for 7.5-20% of the cost. That cost will continue to get pushed down by the Chinese. And bizarrely enough their models are more secure to use because they're open source open weights.
OpenAI and Anthropic are going to get crushed long-term, and their investors are going to take a horrendous haircut.
On the other hand, Google and Microsoft already have the users (and lock-in). They just need to funnel them into Gemini and CoPilot.
DeepSeek r1 affected markets because for a little while people bought this, but it's not true for so many reasons. Sending data to China is out of the question for every American Enterprise. OAI and Anthropic have rich product suites and API harnesses that make DS far from a "drop in replacement." They have better models, generous usage limits, domination of the zeitgeist, integrations with Slack and all the Enterprise SaaS platforms, and magnitudes more GPU capacity than DS. What you say simply isn't true.
- Right, R1 affected markets because the market originally believed your theory, but it doesn't any more, which is why V4 didn't move markets at all.
- Sure, you can use US infra providers. Together.ai is a good US provider but then it's 15X more expensive than DeepSeek's Chinese-subsidized pricing. It's really not that attractive at that price point. Anthropic and OpenAI are focused on larger models, but Grok 4.3[1] is smarter and significantly faster + cheaper than DS4[2] and by a wide margin.
- DeepSeek has a Claude-compatible messages API, but that's trivial. Anthropic has a massive API platform with things like Sessions, Files, and Agents[3]. None of those are available on DeepSeek.
- V4 will definitely move markets, especially as Claude and OpenAI keep jacking up the prices more and more. But inertia exists. Give it time.
- Most US infra providers are ~5x more expensive than Chinese infra, not 15x. But yes you are right. It does erode the cost advantage significantly. Big asterisk is that V4 seems to have solid cache hit percentage, often in the high 90s.
- Grok (and Llama) always underperform relative to their benchmark and ranking results. Don't ask my why, but it's a persistent pattern me and colleagues have noticed. I'll give them another try though, more competition is better.
- DeepSeek themselves have specifically said they prefer developing high performance models that you plug into other tooling, including Claude's. Regardless, I think it's unrealistic to expect DeepSeek to offer 1:1 suite compatibility with Claude or OpenAI. You wouldn't expect that from OpenAI <> Claude either.
I think the real crux of the moat is model intelligence. I'd bet that most of the money being spent on inference is on the top few models (today Opus-4.7 and GPT-5.5) from people and companies that benefit from using the best models.
Truly the main moat that OAI/Anthropic have is being 6 months months ahead of the competition in performance, which might be indefinite if the competition is just distilling their models (China) or takes many months between releases (Google).
Once you look passed the frontier of performance, it's just a race to the bottom on inference costs because there's at least 5 companies with equivalent open models at that level.
I wish DeepSeek were a drop-in replacement, but it's not. It performs amazingly well but it's not as autonomous and needs a lot more nudges compared to Opus4.6/7 or GPT 5.5. It's good enough for a lot of things (text extraction, sentiment analysis, classifying things) but not on the same level for code gen.
It’s still the best option for uptime, document analysis (on a cost basis), and Google is less likely to experience a significant cybersecurity breach than a less established company. They’ll be fine as long as they stay in the game even if they never have a Ferrari again plenty of people buy Toyota.
When will countries start treating cyberattacks as an act of war? If the North Korean military came to America and robbed fort Knox of $200M in gold there would be retribution. But hack an American company for the same amount and the feds do nothing.
Ok, so we treat it as an act of war. Now what? Attack North Korea? Great, the entire city of Seoul gets shelled within five minutes of your attack and hundreds of thousands of innocent people die.
It's very easy to play with lives that aren't yours.
You would be surprised how many people naively think "Why doesn't my country just open a war on X country and this Y problem will be solved forever" in their head they think war is just a flurry of bombardments and the other side (not theirs) is just destroyed to rubble and their country will have only minimal losses
Never retaliating is a great way to get people to attack you. Of course escalating to all-out war provokes the same in response, but there does need to be a proportionate response, because it needs to be stupid to hurt us, not good business. t’s a significant failure of the US government when half the world freely loots US citizens and businesses.
Exactly. This is the "Declare fentanyl a WMD" of solutions to ransomware. Sounds kinda badass as long as you don't spend too long thinking about it but has no practical relevance to actual enforcement challenges.
It's a familiar example of the perennial "[THING] could be solved overnight if [PERSON_OR_GROUP] would just start taking [THING] seriously" trope.
They already do. This is what asymmetric warfare looks like, your weakest links will break in a time of crisis. Focusing on retribution for the Dunder Mifflin cyberattack is pointless, the adversarial motivation is purely to disrupt and extort.
The best response to a cyberattack on critical systems is to take security seriously. Document the offense, avoid the same mistakes and invest in penetration testing. Of course, nobody is incentivized to do that until they're attacked, so the cycle perpetuates itself.
How do you know which country to blame? It is standard practice for foreign actors (or just hackers in general) to use proxies around the world to misdirect and insert false clues as to their origin. It could be an American teenager proxying through North Korea, and it could be a North Korean proxying through another American teenager's residential connection, there's no way to know.
It's touching to hear Anthony Bourdain reminiscing about laundry. The point about a laundry maid giving me back time better spent engineering is a practical dilemma.
I hear you. I also don't buy clothes unless really really needed. But I am hoping that if I do buy a Robot like that, it will do other things as well and not just folding laundry :)
I've never in my life seen a useful product tour. They're always blatantly obvious like "THIS IS THE SEARCH BAR. USE IT TO FIND CONTENT ACROSS OUR PRODUCTS AND SERVICES."
The best UX is using obvious and standard design, plus a searchable menu / command palette.
Ime, the only useful product tours where in games, I. E., tutorials. This usually extends up to in-game hints at certain features like a characters ability. A lot of software can probably pull inspiration from there in regards to including hints with minimal interruption during usage (tooltips that are shown longer the first time you use something etc).
Good question. I find that GPT-5.5 thinking is very good at not thinking for simple questions, so much so that I've never had the need to use the instant model even for quick Q&A.
I'm assuming the instant model, then, is an entirely different smaller model mainly serving the free tier of ChatGPT.
Good point. I feel like this does a disservice to ChatGPT -- IIRC even the free tier of Claude points you to Sonnet 4.6 by default, which is magnitudes better than 5.3-instant which has been the default in ChatGPT.
Hence most users will immediately think Claude is smarter, even if their best models are on par.
Correct. I have the $20/month plan and I just checked, the default is 5.3-instant. I can manually switch it to Thinking is 5.5. I also have it set to auto-switch.
Counterparty risk would be more like the betting site going belly-up and not getting money out despite having a profitable trade. This is just regular risk.
This actually isn't the case anymore, plenty of non-US airlines allow beards and the new masks still provide perfect oxygen flow. It's more of a "professional image" thing in the US.
reply