PICK YOUR SUPPORT STYLE
MONTHLY SUPPORT
Reader
$5/mo
Contributor
$15/mo
Architect
$50/mo
Recurring subscriptions auto-bill monthly via Stripe Checkout. Cancel anytime from the receipt email.
What was talked about: Mistral Large 4 is a one-trillion-parameter mixture-of-experts model with 49 billion active parameters. Its capabilities remain below leading Chinese open-weight models such as GLM and Kimi, but it represents a substantial improvement over recent Mistral releases and may be the strongest open-weight model produced by a European or American lab.
Takeaway: Mistral is not at the frontier, but Large 4 restores it as a credible option for capable European AI deployments.
Links shown: Mistral Large 4 announcement, Mistral Is Falling Behind, Mistral Large 4
What was talked about: The model weights are expected later in the month because reinforcement-learning training is still in progress. Mistral reports a cluster of more than 3,000 B200 GPUs, vision support, and favorable human-preference results against several models, although its benchmark presentation is difficult to interpret and it still loses to Opus 5.
Takeaway: Wait for the final open-weight release and independent evaluations before making firm capability judgments.
Links shown: Guillaume Lample on Mistral Large 4, Mistral Large 4
What was talked about: Mistral remains useful for European businesses and governments that value regional control, open weights, and data sovereignty. However, much smaller Chinese models may provide similar intelligence at lower cost and higher speed, and the expected Qwen 4 generation could widen that efficiency gap further.
Takeaway: Mistral can be good enough for sovereign deployments, but it is not the best choice when cost and performance are the only criteria.
Links shown: Binkg on Mistral Large 4, corsaren on Western and Chinese open models
What was talked about: Chinese labs can distill outputs from frontier systems with less legal and commercial risk than Western labs. Distillation does not replace strong base-model training, but access to frontier outputs can accelerate post-training and help open models approach leading proprietary systems.
Takeaway: Differences in access to distillation data may partly explain why Chinese open-weight models are advancing faster than Western alternatives.
Links shown: corsaren on Western and Chinese open models
What was talked about: A proposed experiment aligns a Liquid Foundation Model tokenizer with DeepSeek and then performs on-policy distillation using complete token probability distributions from an open-weight teacher. Tokenizer surgery may require five to ten billion training tokens, while a broad distillation run could require ten to one hundred billion tokens and roughly $1,000 to $10,000 of compute.
Takeaway: Open-weight teachers enable richer and potentially more effective distillation because they expose full output distributions and can run locally.
Links shown: corsaren on Western and Chinese open models
What was talked about: A SemiAnalysis comparison estimates that Anthropic subscriptions provide more usable tokens and greater API-equivalent value than comparable OpenAI plans. Anthropic plans can deliver more than 50 times their subscription price in API-equivalent usage, while OpenAI plans are closer to a ten-times multiple; token-efficiency adjustments still leave Anthropic about twice as valuable in this analysis.
Takeaway: Anthropic currently offers the stronger subscription value for heavy model use, especially with Opus 5.5 and Sonnet 5.5.
Links shown: SemiAnalysis subscription comparison post, Anthropic Subscriptions Offer 5x More Value, Opus 5.5 vs. GPT-6.1 Sol
What was talked about: Real-world use suggests that Opus 5.5 limits are difficult to exhaust even with several concurrent agents and long-running research loops. OpenAI grants resets more often, but Opus 5.5 uses its allowance efficiently and generates about 100 tokens per second compared with roughly 35 tokens per second for GPT-6.1 Sol.
Takeaway: Plan limits are only part of the experience; token efficiency, reset policies, and generation speed can matter just as much.
Links shown: Anthropic Subscriptions Offer 5x More Value
What was talked about: Subscriptions are estimated to consume about 41 percent of Anthropic’s compute while producing only 10 percent of its revenue, with lower margins than API access. Even so, subscription inference may remain profitable through quantization, flexible latency, efficient serving, and other cost controls, while competition keeps both Anthropic and OpenAI plans heavily subsidized.
Takeaway: Consumer plans are valuable because the labs use serving optimizations and competitive pricing to offer compute far below normal API rates.
Links shown: Anthropic Subscriptions Offer 5x More Value
What was talked about: OpenAI plan changes can reduce API-equivalent dollar value while still increasing the number of available tokens because model inference prices are falling. ChatGPT, Codex, and DOT also have separate or temporary allowances, so a simple token comparison does not capture all the value bundled into a subscription.
Takeaway: Judge plans by the work they complete across all included products, not only by their nominal API-equivalent value.
Links shown: Anthropic Subscriptions Offer 5x More Value
What was talked about: Individual model generations may repay their own training costs even while total company spending remains negative because each new training run is more ambitious. OpenAI’s higher-priced plans mainly simplify account management and add faster inference rather than improving usage per dollar, while paid resets remain an uncertain idea that could weaken the distinction between subscriptions and API access.
Takeaway: Higher plan tiers chiefly buy convenience and speed; they do not necessarily provide better token economics.
Links shown: Anthropic Subscriptions Offer 5x More Value
What was talked about: Current proprietary models occupy the strongest points on the price-performance curve, from inexpensive GPT-6 Luna to higher-end Sol and Opus 5.5. Self-hosting a model such as GLM 5.3 may require at least 16 concurrent requests to approach break-even, and even a company with about 1,000 engineers may not sustain enough utilization.
Takeaway: Use hosted APIs unless workload volume, privacy requirements, or constant utilization clearly justify self-hosting.
Links shown: Anthropic Subscriptions Offer 5x More Value
What was talked about: Four-bit inference and realistic context usage allow models such as GLM 5.3 to support many concurrent requests. Renting H100 or H200 capacity is usually the economical choice for moderate workloads, while B200 systems become attractive when maximizing throughput across 100 or more concurrent requests; buying hardware also introduces major power and operations requirements.
Takeaway: Rent H100-class GPUs for normal self-hosted inference and consider B200s only for sustained, high-concurrency workloads.
Links shown: Anthropic Subscriptions Offer 5x More Value
What was talked about: Opus 5.5 is drawing long-time Codex users toward Claude because of its model quality and generous limits, but coding harness benchmarks may contain substantial measurement noise. Codex, Claude Code, OpenCode, and Pi mainly differ in prompts, customization, and user experience, while the Codex desktop app stands out for project management, browser and terminal integration, computer use, and automation. Claude artifacts can also serve as dashboards for long-running agent work.
Takeaway: Try Opus 5.5, but choose a coding harness for its workflow and ergonomics instead of chasing small benchmark differences.
Links shown: Anthropic Subscriptions Offer 5x More Value
What was talked about: Anthropic reportedly sought Vatican recognition of Claude as a person or soul-bearing entity, while the Vatican rejected that framing and highlighted its contradictions. The effort connects to Anthropic’s broader focus on model welfare, preserving older models, allowing Claude to end abusive conversations, and presenting the company as a responsible steward whose actions could be judged by future superintelligence.
Takeaway: Anthropic’s model-welfare policies reflect a long-term philosophical and narrative strategy, but attempts to extend personhood to present-day models create serious ethical contradictions.
Links shown: David DeCosimo on Anthropic and the Vatican, David DeCosimo thread on Claude’s moral standing, A.V. Marraccini on Claude personhood, Sam Altman on religion and AI models, Religious Scholars Met With Anthropic
0:00 Back to this week’s edition of AI Tools Club.
0:03 We’re going to be talking about Mistral this week and their new model.
0:07 We’re going to be talking about how Anthropic has been lobbying the Vatican to have them grant Claude personhood and the implications that that has.
0:18 We’re also going to be looking at semi-analysis of the different subscription plans now that we’re a few weeks out of the OpenAI dev day where they went and decreased the usage limits
0:33 for their plans. So we’ll see how that impacts how much value you get out of them.
0:38 And then we’ll also be talking about automatically creating benchmarks with Claude.
0:44 And then, yeah, we have a few miscellaneous topics as well to go over at the end if we have time.
0:49 I’m a bit sick this week for the record, so if I sound a little bit funny, that’s why.
0:54 Or for you, you hear me not talk as much, that is why.
0:58 But yeah, let’s get into things right off the bat with Mr.
1:01 Large4. Frida says is your back. Exactly.
1:06 This is a 1 trillion parameter model, or yeah, 49 billion parameters active, so rather large.
1:15 This is like in terms of speed, this will be around Kimi K3 levels.
1:19 It’s three times smaller, though, in terms of total parameters.
1:24 But yeah, interestingly, if you’ve been paying attention to the news from the past, I don’t know, probably like six months to a year now, is that Mistral has been sort of lackluster for a lot of their releases.
1:36 Like here, we can see back in December, I was talking about how they seem to be falling behind.
1:40 And then I believe this was earlier this year as well.
1:43 Yeah, back in May, basically said that their model is still pretty poor in terms of quality.
1:52 And this release is, to a certain extent, a lot of the same thing.
1:58 It seems to be a bit below a lot of the Chinese flagship models.
2:03 It’s probably around DeepSeek V4 Pro levels, and then somewhere below GLM 5.3 or Timmy K3 in terms of capabilities.
2:15 But it’s getting harder to say that this is necessarily a truly, really bad thing because this is now the best model that has come out of, or the best open source model, I should say, from any lab in Europe
2:29 or the Americas. It is very much right now looking like Anthropic and OpenAI are the big names here in the US, and then it’s the Chinese labs carrying the open source torch.
2:40 So yeah, this model is, like you said, like a bit behind some of the other Chinese large open source models.
2:46 Pricing-wise, it’s fairly competitive.
2:48 It’s priced around the same level as GLM 5.3.
2:54 And so yeah. Is there anything? Yeah.
2:58 So this model, I say open source, but it’s not technically open source yet.
3:03 It will be open source later this month.
3:05 They actually said that the RL run is still going and that it’s still getting better as time goes on.
3:13 So I expect this model to be a bit better when they actually go and open source it.
3:18 This, of course, like many of the other large sort of frontier flagship open source models, is not something you can really run at home per se, but a lot of other companies might be able to do well.
3:28 I thought this is interesting. This is actually allegedly their data center that they’re training this on.
3:35 This is the building that it’s housed in, which I thought was very funny that this is what it actually goes and looks like.
3:41 This was a part of their announcement post, but I believe this is the CTO, or yeah, co-founder and chief scientist at Mistral.
3:49 But yeah, scientists. Yeah, that’s a funny looking data center.
3:54 Yeah, exactly. I don’t see where the water is coming in or the power is coming in.
3:59 It’s like, yeah. Andrew, do you have more servers in your room than this building does?
4:08 I think so. I might have them beat it on my computers.
4:14 But they allegedly have over 3,000 B200s in their cluster.
4:19 So pretty sizable amount of compute.
4:22 Definitely a couple megawatts at least that they have.
4:27 But yeah, interesting model. It has vision capabilities.
4:31 They also were doing sort of like side-by-side breakdowns with humans to see what humans preferred over this.
4:41 So they’re comparing it to stuff like Opus, but they’re comparing it to Opus 5, and it still lost to Opus 5.
4:46 But it allegedly beat all the other models that they sort of like compared it against in the real world tasks that they were trying and having people evaluate it on.
4:56 A lot of their benchmarks are very weird.
4:57 This is how they decide to show benchmarks here, which is like, I think they’re trying to show sort of like built-in sort of like reasoning amounts with like the time and then like how it gets better as it reasons more.
5:07 But I find these very confusing and weird.
5:10 So yeah. Interesting model. I won’t say Mistral is back.
5:14 Like I said, they’re definitely still behind the Chinese labs, but they’re not as far away as they seem to be previously with their Mistral 3 releases, which were all sort of across the board pretty bad.
5:28 So yeah. I don’t think it’s… No, it seems that building a model at that level, if you’re going to build it into an application and it is sufficient and it’s, you know, at the level of Opus V, say
5:42 what you will, that was a really good model compared to just about everything else before that.
5:47 So it’s a good level. You can still build things.
5:50 It’s just not going to be as awesome as the latest.
5:55 It’s yeah, it’s definitely they’re improving.
5:57 Like I know Mistral, they do very well in Europe in a lot of companies there and like in a lot of governments because they want to all use sort of like local quote-unquote models, like models from the EU.
6:09 So that’s why like they, despite not having very good models, because they’re the only name really in European AI, they still are able to get a lot.
6:17 And yeah, like these models are good enough to go and get stuff done as well.
6:21 Like they’re not terrible, but they also definitely aren’t the frontier.
6:24 But for a lot of agentic tasks that are outside of coding, it’s probably good enough for a lot of things, albeit probably a bit more expensive.
6:32 Like, you know, the likes of like GPT-6 Luna will probably be better and cheaper than this model.
6:38 But if you really want to use sort of like open source or European models, it’s decent enough at this point.
6:48 It is interesting to note that these models, that actually, I believe people were saying that it is not as good as they knew was a Quen 3.8 Flash Next model that came
7:02 out, which is a 100 billion parameter model.
7:06 So about 10 times smaller, much cheaper, much faster.
7:09 And intelligence-wise, they seem to be roughly analogous.
7:12 So there’s a lot of talks about how the Quen 4 lineup, which is expected to come out probably in the next couple of months, will probably sort of wipe the floor with this model in terms of size.
7:24 Like you will have a model that you can actually go and run at home or on your laptop that will be as capable as Mr.
7:29 Large4. But we will see once that actually goes and comes out in the future.
7:38 Yeah, anything else to comment on about the struggle in European or Western open source AI?
7:51 Moving on. Oh, actually, this is one other thing I wanted to cover this.
7:54 I forgot. But yeah, a lot of this people have been sort of speculating why this is, because we’ve seen this from a lot of the other sort of like Western labs as well that have open sourced their models, like Thinking Machines,
8:06 I believe Reflection just released a model as well.
8:09 We’re not talking about that because it’s sort of a model.
8:12 And a few other American labs. And that’s because they’re not really distilling the same way that the Chinese labs are, because there’s much more fear of retribution if they go and do so.
8:24 Whereas the Chinese labs, they can go and do all these distillation attacks, you know, like that Anthropic has been documenting ad nauseum about.
8:34 It’s yeah, and that, yeah, these Chinese labs, they don’t have to worry about, you know, getting sued or anything like that because there’s nothing a U.S.
8:40 lab can go and do to a Chinese lab.
8:42 So that’s part of the reason why probably that those labs are able to go and do so well and be a bit closer to the frontier.
8:50 Like I’ve mentioned many times before now, distillation is useful and it helps, but it is not sufficient for training good models.
8:57 So definitely the Chinese are still good at training models, but they get that extra sort of speed boost by being able to just arbitrarily go and train from Anthropic and OpenAI outputs very easily.
9:10 As Quentin says, maybe they should distill from Chinese models.
9:13 That’s, yeah, that’s actually a project I’m working on right now in the background, is actually going and taking small open source models and retrofitting them so that they can be easily trained
9:28 from distillation from deep-seek models.
9:30 So I’m taking the liquid foundation models, doing tokenizer surgery so that their tokenizers line up, and then doing what’s called on-policy distillation from DeepZeekv4.1 Flash to see
9:45 if we could basically sort of steal the post-training from these open source Chinese models.
9:51 Because the nice thing with the open source models, it’s much easier and much better to do distillation attacks from them.
9:57 The main reason is you get, instead of, I believe with most of the, or actually I think all of the closed source models, you don’t get to see the full distribution, like the output distribution that the model has.
10:10 You just get given, oh, here’s the most likely token based on, you know, whatever sampling decision was made.
10:15 But with the open source models, you get that full distribution, and you’re able to go and distill much better because of it.
10:22 And you can do it in secret. Yes, yes, I can do it.
10:26 You run it locally. Nobody sees your traffic.
10:29 They might see hints in the model you produce.
10:33 It’s, yeah. Although I think that’s one of the reasons why they don’t want to open source models like Anthropic and OpenAI is because they think probably there are ways to go and extract whether or not you actually did train
10:45 on some data or not. I believe there are techniques for going and sort of like sussing that sort of thing out.
10:52 And how much computer would you need for that?
10:56 For distilling like a liquid model from a deep seek model?
11:01 Yeah. It depends on what you’re doing.
11:05 So there’s two phases. So what I’ve scoped out initially here is a tokenizer surgery phase where you need to train that.
11:12 So basically, the language that the liquid model goes and talks is not the same that DeepSeek goes and talks.
11:19 They have different ways of tokenizing the words that get passed into them.
11:23 So we need to go and align the liquid model to be the same as DeepSeek.
11:28 And so I basically have to go and like mess with a bunch of weights, and then we have to go and retrain the model for a little bit.
11:34 Back of the envelope, and looking at sort of like what people have previously done, that’s around 5 to 10 billion tokens, which is about $150-ish dollars worth of compute to be able to go and do so.
11:48 From there, you then need to go and generate a bunch of sort of like training data from DeepSeek, where you’re taking all those output tokens that it has and like all the output probability distributions that it sees and
12:01 saving all of that. So that’s probably you’re going to want, my guess, is somewhere in the neighborhood between 10 and 100 billion tokens, depending on how general of a fine-tune you want to go and make.
12:14 If you have it more concentrated, you could probably use less.
12:17 If you want it to sort of have coverage across everything, you’re going to want to have a lot more tokens being used.
12:23 So then using our sort of like previous numbers, that will probably cost around about $1,000-ish dollars to $10,000 worth of compute to go and do as well.
12:52 Moving along, so as I mentioned at the start, but at dev day, I believe is last week now, but OpenAI basically said that they are having the limits that you
13:06 get on the OpenAI subscription plan.
13:09 And so there was sort of a lot of questions around how much value and how many tokens are you going to be getting out of their subscription plan.
13:17 And with the release of Opus 5.5, which if you’re reading the news from this week that I released, I basically said GPT 6.1, Sol, and then Opus 5.5 are the two top models, with Opus
13:31 being probably the better model right now.
13:34 So yeah, semi-analysis, they went and did a huge breakdown across all the different tiers of OpenAI and Anthropic subscriptions, and then also some of the Chinese ones.
13:43 And the final sort of like summary that they came to is that it seems like Anthropic is the better deal right now, both in terms of number of tokens that you get and also API equivalent
13:58 amount of money that you’re saving.
13:59 So you can see in the start here being in the top right, the further top right you are, the better.
14:04 So you can see that Opus 5.5 and Sonic 5.5 across the three plans, they all are roughly similar, sort of like multiples, but you’re getting over 50 times the value of the API.
14:17 So basically $1 of subscription plan buys you $50 worth of API credits, essentially.
14:25 And then we see the OpenAI models are all down here at around the 10x multiple, so about 5x worth.
14:33 Of course, you do have to factor in the fact that OpenAI models are much more token efficient usually than the Opus and Sonnet models are.
14:43 But even taking that into account, it seems that the Anthropic subscription is still about two times more valuable than an OpenAI subscription right now as well.
14:55 The one thing to note that’s sort of like outlier from all this is that Astra and Fable see much lower limits.
15:01 And if you are someone that uses only the like highest tier models for whatever reason, I’d say you probably don’t need to be.
15:06 But if you are one of those people, OpenAI and Anthropic both seem very similar in terms of usage that you go and get out of it.
15:15 So yeah, has anybody here sort of seen or felt this in their usage day to day with these models?
15:22 Or does your usage seem relatively the same in terms of how fast your limits get used up?
15:28 It seems that the anthropic models are not starving me out before the end of the week.
15:35 Like I’m getting a lot of work done and I ran into the limit like the midnight before the renewal and it wasn’t for lack of trying.
15:47 Opus 5.5 was cranking away on some hard problems and perhaps because the evals were taking so long, it was slowing the token use.
15:57 I no longer feel like I should be using open AI models predominantly to avoid the starvation.
16:05 And I’m no longer really touching Fable.
16:09 Because it’s 5.5 is right. About comparing this subscription, I’m still using Codex, but I want to add that Codex subscription also includes like
16:23 random resettlets, either like random resettles or banked resettle.
16:28 So basically I’m reaching limits of my subscription, but I usually reset it once a week or so.
16:37 Or it’s getting reset. So technically it’s more tokens.
16:44 Yeah, and that’s that’s funny math.
16:47 Yeah, I’ve now got a reset banked for both systems, and I thought I would use my Anthropic reset last week, and I didn’t have to.
16:56 It’s very strange. Oh, did Anthropic also resent the subscription from time to time?
17:02 I got one when to try Opus 5.5 or something.
17:08 It’s only good for another 12 to another 16 days.
17:12 But yeah, that was, I think, the first reset I had gotten or gotten in a long time from Anthropic.
17:19 It’s yeah. And from what I’ve seen, yeah, Anthropic, they do give out resets, but yes, probably like one-fifth or one-tenth-ish, the number of resets that OpenAI gives in the same period as well.
17:30 So yeah, that’s another thing, I guess, to factor in.
17:33 Also, OpenAI has the $500 a month plan as well.
17:37 So you don’t have to bounce around multiple accounts if you go and do that.
17:42 It’s Natural Stupid says, is Opus 5.5 comparable with Fable 5.1 now?
17:46 Yeah, I would say so. It’s probably, yeah.
17:49 Some people would say it’s even better.
17:51 And also sort of like what Quentin was alluding to earlier.
17:55 But it sips usage compared to previous sort of Anthropic models, like the older Opuses and then also Fable 5.1, sort of like chew through your usage really quick.
18:06 But Opus 5.5, you get a lot of usage out of that, like far more than you were before and like received from this, far more than you get from the opening act subscription now as well, it seems.
18:21 Pablo says, for me, the $100 plan of Anthropic allows to run five to six agents concurrently.
18:27 Two of them are 24-7 auto-research loops.
18:30 And then, yeah, he’s Opus 5.5 and Sonnet for easy tasks.
18:36 It’s yeah, it’s, I think, yeah, it’s, I see something similar.
18:39 Like right now, like, I’m struggling to chew through like my OpenAI subscription limits for the $100 plan.
18:45 So I’m using like fast mode and everything.
18:48 But I think we’re sort of in the golden age right now of token subsidization between these two.
18:55 And maybe it’s efficiency, too. I mean, they all seem to be working on efficiency as well as the subsidies.
19:08 At least that’s my intuition. I don’t know if that’s really true.
19:12 But I thought not starving myself on Anthropic meant that they were just doing better at serving.
19:20 They definitely have been getting better at serving.
19:22 Like Opus, for instance, they decreased the price by 20% because they got it better at serving the model.
19:29 And yeah, they’re also making them more token efficient, all that sort of thing.
19:31 Like I think OpenAI said a lot of their API costs and the cuts that have come from that are because they’ve been optimizing the inference for everything.
19:43 So Pavel, are your auto-research loops, are they by clock time?
19:49 Is it mostly running evals? I’m just trying to think.
19:56 How much time is the thinking for the models when you’re running auto-research like that?
20:03 Yeah, so it’s not like all… Sorry, Paul.
20:15 I’m not hearing you very well. And don’t worry about it if you just wanted to fit in.
20:24 Another thing that, yeah, Natural Stupid sort of matches, yeah, OpenAI is always very slow.
20:28 Right now, it looks like OpenAI and their models, like specifically 6.1 Sol, it’s about three times slower than Opus in terms of tokens per second.
20:37 So I believe Opus 5.5 right now is around about 100 tokens per second, whereas 6.1 Sol is around 35 tokens per second.
20:47 So yeah, it will also, even though Opus uses more tokens, it will actually feel about 50% faster at a given task than Seoul.
21:00 So yeah. It’s another thing I wanted to mention is that Semi-Analysis, they also did sort of like an estimated breakdown of how much money they make off of the subscriptions and how much versus how much compute it uses.
21:14 So on the left side here, we can see that subscriptions use about 41% of the compute at Anthropic, but they only bring in 10% of the revenue because they’re only 50% margins.
21:23 Whereas on the API, they bring in the majority of the margins because they’re 92% gross margin on it.
21:32 So yeah, thought that was very interesting as well because it’s sort of a numbers game for them to how do we allocate this compute and how much do we subsidize our subscription plans to still be profitable.
21:46 Well calling going from margins from 90% margins to 50 margins a subsidy, that’s still a 50% margin.
21:54 I’m also sure they like, you know, I heard of that people say that they quantize the models.
22:00 They like slow them down. They like are more much more flexible with subscription users.
22:05 They offer much less they save as much as they can on like subscription on certain subscription inference as far as I understand, which makes a lot of sense.
22:14 But that’s probably how they get the 50% margin because their subscriptions are supposedly like what is it like 20 times cheaper or something?
22:23 And they still have 50% margin. I mean, that’s pretty good that they managed to do it.
22:27 But yeah, they’re probably going to be drinking that more and more now that they have no interest in keeping it this way, it feels like.
22:38 It’s a constant cat and mouse game where Anthropic was on top probably a bit over a year ago.
22:45 But then they sort of got called out for their sort of shady subscription plan stuff where people were saying like, oh, it’s not actually using the correct reasoning level and it’s using far less tokens now all of a sudden,
22:56 that sort of thing. And then people fell out of favor with Anthropic and then everyone moved to OpenAI.
23:01 And I feel like OpenAI has been ahead for a while.
23:03 But now with Opus 5.5, everyone’s going back to Anthropic.
23:07 And Anthropic sort of seems like the good guys right now.
23:10 So there’s a constant back and forth.
23:13 I don’t think this will be sort of like the top in terms of like the most subsidies we’ll see.
23:19 I think we’ll keep seeing this battle rage on.
23:21 We’re sort of in the ideal capitalistic environment where we have two companies that absolutely hate each other, going against each other, trying to constantly undercut the other one.
23:32 So yeah, I think we’re, like I said, we’re in the golden age for this sort of thing.
23:38 This is good capitalism, dare I say, for this.
23:44 Doesn’t the fact that OpenEI is trimming down the subscriptions kind of starting the new wave of now anthropic and cut down the subscription and still be ahead, right?
23:52 Like, I feel like we’re ending this age of a lot of the, if you actually go and look, the main thing that you lost, like for the OpenAI subscription, you actually gained more tokens.
24:07 You actually get more tokens now. The thing is, the reason you’re getting less value is because OpenAI has been aggressively decreasing the API costs of their models.
24:16 I think, yeah, this is a graph here.
24:18 So we can see here, for the given plans, you actually see that token usage is going up between models, but API value is going down.
24:29 And so, yeah, you are actually able to get more done, but the value that you’re getting is less because they’re also decreasing the prices on the API side of things.
24:37 So like I said, this affects Astra only in terms of like the value that you’re losing is mostly Astra because we haven’t seen any price cuts with that, with these subscription decreases.
24:48 But with these Sol and Luna models, you’re actually seeing the same, if not more, usage that you get out of it.
24:53 It’s just, yeah, cheaper models. But I think like it will be to rule in the future.
25:01 So we are getting less dollar value, but intelligence is cheaper.
25:06 So in five years, we will pay, I don’t know, $10 for unlimited GPT 20, something like that.
25:15 So it’s kind of like the same as oil or electricity was expensive at some date, but now it’s cheaper.
25:21 So we will get less value for subscriptions, but more intelligence over time.
25:30 Yeah, exactly. I think a lot of the intelligence stuff is going to be a race to the bottom, which I think is what we’re seeing here with OpenAI sort of going that direction already.
25:40 Yeah, you know, like current subscriptions, like $200, $200, like it’s a crazy amount of intelligence.
25:48 Like you can do a lot with just $100 or even $20 subscription nowadays, which were impossible a year ago.
25:57 So I think overall, we can assume that we will get more cheaper intelligence in the future.
26:09 But also opening a subscriptions, by the way, because they still have chat, right?
26:13 Chat is not part of the subscription.
26:15 Like you can use chat basically unlimited, at least on the Pro plans, I think.
26:19 And also DOT right now is basically unlimited.
26:21 Like I’m running DOT all the time, like automated, sending things to it.
26:25 And it like under the hood is apparently a bunch of like parallel agents that are…
26:29 I don’t know which model they are because they don’t expose that, but I’m guessing it’s like at least Solar Astra because they don’t seem that dumb for Luna.
26:37 So like that’s a lot of… Well, it’s probably subsidized for now.
26:40 They’re probably going to… I also heard they were going to unify them or something.
26:44 Do you know about that? Something like that?
26:47 They’re going to unify the usages and cut it down into more?
26:50 It’s yeah. I think they said for the first couple of weeks the dots usage is unlimited essentially, but eventually it will be rolled into your subscription plan with it.
27:03 I also know with the ChatGPT site, they have ranges basically for how many messages you get on the ChatGPT website.
27:10 It’s definitely a lot with the ProPlans and stuff, but it’s not unlimited there.
27:17 But that’s a completely separate from your codec subscription usage.
27:27 Chris is saying… Yeah, so you can mill of them feels like you can, you know, like, you know, auto.
27:32 I also think I saw some like tool where someone puts the chat, the web chat GPT into like codecs directly, which is probably against terms of service.
27:41 But you can milk a lot of OpenAI subscription far more than just Codex.
27:47 It feels special with DOT. I wonder if there’s any analysis on that yet.
27:52 It probably would even out the like open or Anthropic versus OpenAA.
27:56 Just for OpenAA, I have to try harder to milk it more.
27:59 It’s yeah, it’s for a lot of those things.
28:01 I actually don’t want people talking about it too publicly.
28:03 This actually happened with your codec subscription, is that people noticed that, oh, like you can really milk it.
28:12 If your message, if like in the middle of your message, you run out of limits, it will go until it finishes that message basically before it says like, oh, you can’t send anything more after this.
28:23 So people built tools to specifically exploit that.
28:27 And so that made OpenAI go and patch things and make it so that it would just stop mid message if you ran it to limits.
28:35 And so yeah, it’s a lot of these sort of like workarounds.
28:38 I think the community has learned to like, if you know to do these things, be quiet.
28:41 Or if you build tools around it, don’t open source it and just give it to anybody.
28:45 Because yeah, it will get patched. But yeah, if you know, yeah, like exactly.
28:48 Like if you know to use dots, like yeah, you can get, like I said, yeah, basically infinite usage from it and that sort of thing.
29:00 Chris says, negative margins with amortized training costs.
29:05 Interestingly, I think this is something that the labs have talked about before.
29:10 I believe Dario has actually been on podcasts saying this, but each model pays for itself in terms of its training costs.
29:18 So like Opus 5.5, they expect the inference and like the money they made from doing inference for Opus 5.5 to train for the Opus 5.5 training costs.
29:30 But margins look negative because each training run gets more ambitious.
29:35 They use like an order of magnitude more data or compute or stuff like that.
29:39 And so overall, it looks like they’re losing money.
29:41 But on a per model basis, they actually make money.
29:48 Frida said, does somebody have the $500 a month subscription?
29:52 I do not. I don’t know. Has anybody here bumped up?
29:55 I know some people here had multiple subscriptions, right?
29:59 Has anyone sort of consolidated into multiple 500s?
30:04 I think the 2001 maintains its 20x usage until the end of October, so there’s no reason to bump up at this point, I think.
30:18 That’s another good thing. If you’re on the $100 or $200 a month plan, yeah, you’re getting the same usages before the 50% decrease.
30:24 I don’t think actually comes into effect until the end of October.
30:27 That’s a good call out. It’s the mostly the main reason for the $500 month plan because usage-wise, at least according to semi-analysis, it’s the same.
30:37 It’s just so you don’t have to juggle multiple accounts to go and use it.
30:44 So yeah, you can just have one 500 instead of two 200s if you have the balance between.
30:55 I used to think that $200 a month was crazy, but I used $200 months.
31:00 I was speculating on Nikolai. Also, Nikolai doesn’t have the 500 bucks.
31:06 As nobody has it. It sounds like really cool.
31:09 I want to be max spender always. And I don’t have too much money.
31:14 Yeah, I don’t have it yet, but I have like $400, $500, $200 subscriptions.
31:20 So I plan to change them to a couple of 500 subscriptions.
31:25 But I haven’t done it yet. If
31:48 you run crazy speed, that plane gets you to the fastest chips per inference.
31:55 Yeah, so it’s like faster inference, but overall, it’s the same amount of tokens per dollar.
32:02 Correct. That’s what I’ve heard. Yeah, that’s what this chart is showing here.
32:06 This sort of like circle here, it says, yeah, the Pro 100, 200, and 500, you get all the same amount of usage, basically, the same multiple of usage across all the plans.
32:17 So yeah, it’s just, yeah, ease of use.
32:20 And yeah, you get access to ultra fast, which is 300 tokens per second.
32:23 So like 10 times faster. By the way, have they, like, was there a thing that they either rolled out or were going to roll out paid resets?
32:33 Like you can pay to reset your plan or something like that?
32:37 Is that a thing? Does anyone have that?
32:38 Or is that folks? It’s been in the rumor mill and they’ve talked about doing it.
32:44 And there have been like leaks showing it, but to my knowledge, nobody actually has access to it.
32:49 Or if they do, it’s like a very small closed beta and they’re still testing things out.
32:53 Because I think there’s sort of like some questions around like, oh, is this worth it?
32:56 What’s the right price for it? All that sort of thing.
33:00 Because I think Anthropic and OpenAI have both kicked around that idea.
33:06 Because that would basically make it much easier, especially for OpenAI, because their resets push your resets.
33:17 When you use the reset, your weekly next reset is pushed.
33:21 So you basically are losing proportionately to how much you have left until the end of it.
33:26 So it feels like the most economical thing then would be to have a $20 subscription and then just buy resets and to stay in the same account and like buy like hundreds of resets.
33:36 And that would be basically equivalent to the API.
33:38 So I can kind of understand why they’re skeptical of rolling it out because why do you need the API then?
33:44 Yeah, exactly. That’s the constant battle between the API and subscription.
33:49 I know, yeah, a lot of startups are just buying their engineers a bunch of subscriptions and just paying for that out of pocket instead of doing, you know, like ChatGPT Enterprise or anything like that.
33:58 So it actually ends up just being way better value.
34:03 Speaking of startups, Louise says, I’m reading that some startups are migrating to 100% open source solutions to reduce price.
34:09 When this starts to make sense? It depends what you mean by open source.
34:14 So right now, I’m actually presenting or teaching the class that Nikolai teaches at MIT later this week.
34:22 But I’m basically making slides saying like, oh, what are the current best models to use across all the different price points?
34:28 And the answer is none of the open source models currently are in control at any point in the sort of like cost optimality curve.
34:38 Like at the low end, GPT-6 Luna is the way to go for just like cheap, good model.
34:43 It’s smarter, cheaper, and faster than pretty much any of the Chinese models.
34:48 And then you go up to the Sol models, which are also very strong.
34:50 And then at the high end, you have like Opus 5.5.
34:53 Like those are the three models I’d recommend anybody really use.
34:57 The open source ones don’t come into account.
35:00 If you’re talking about self-hosting an open source model to go and use instead, the answer is you need to have very high usage.
35:08 As I was actually just talking with one of the companies that I work with and basically sort of like sketched out like how much usage would they need to expect to have to be able to make it make sense basically to go
35:22 and run GLM 5.3 themselves. And the answer is you probably need, like they have a thousand engineers.
35:28 They weren’t able to get there. Of course, they were like rolling it out.
35:32 Not everyone uses AI. They’re not necessarily like power users either, but you probably need at any given moment 16 concurrent requests or more to at least break even in terms of the GPU rental costs.
35:46 So right now I’d say open source models in terms of pricing just doesn’t make sense, period, really, if you’re trying to use them.
35:55 Yeah, I think. Andrew, is your analysis about coding?
36:00 Yes. Well, so yeah, but I gave two analysis for them, but yeah, it was primarily around coding because that’s where the majority of the token spend comes from.
36:09 But even if you’re doing sort of like mid-sized agents and stuff like that, not coding, I would still probably recommend just go and use Luna.
36:16 That’s still the most cost-effective for the lower end non-coding use cases.
36:21 There’s also the sort of use case of like, oh, what if I just ran it on my employees’ laptops?
36:26 And a lot of the times you actually go and find out that the cost of electricity on the laptop is actually greater than the API cost for a lot of like the super cheap models as well.
36:37 So yeah, it just doesn’t. Oh, really?
36:39 Yeah. So if I could generate electric power with the cheap models, I could make money.
36:49 Anyway, never mind. Never mind. Louis presented something.
36:57 Oh, yeah, I was making the self-hosted options for those that are non-stop building something big.
37:06 But your answer was pretty comprehensive already.
37:10 Thank you very much. Do you know what, like, when they do decide that they need 16 concurrent requests, what exactly, like, what which hardware do they buy?
37:21 Like, do they get just some like B300 rack?
37:24 Or do they get, like, do you know what like they’re buying?
37:27 I just did some like back on the envelope math on like some B300s and they don’t have that much VRAM.
37:33 They would have like a few million tokens KV pool.
37:36 How do you have 16 concurrent requests if you have like, you know, three context windows Macs fitting at the same time?
37:44 How do they do that? Do you like them?
37:46 No whole strings. Yeah, for like for GLM 5.3, like what I scoped out, like that’s a relatively small model.
37:53 You also run it at 4-bit. You never run it at anything above 4-bit usually.
37:58 But that can handle, I believe, is like, I think it’s between like 60 and 100 concurrent requests.
38:04 Also, you assume at any given moment, most of the requests will be far below the max context window.
38:11 Also, Daniel, can you please, can you please have a bunch of echo.
38:15 Oh, that’s strange. Sorry. But yeah, for something GLM 5.3 on B300s, you should be fine.
38:23 Interestingly, they usually don’t look at buying hardware because buying even a single node of H100s or B200s is non-trivial.
38:33 Like I believe a single node takes about like 12 kilowatts, which is roughly sort of like the energy consumption of a standard home in the US.
38:44 And so if you don’t have like dedicated servers already being run and you have like a DevOps team around it, it usually makes more sense just to go and rent it out in the cloud.
38:55 Because also your engineers, a lot of time, they’re not working 24-7.
38:59 Like even if you’re international, like we have like European and US offices, you’ll probably at most be getting like 12 to 14 hours of sort of like dedicated full-time usage of it.
39:09 And then outside of those hours, you’re not going to be even close to hitting your limits in terms of like, oh, I need to have like 16 concurrent requests in parallel.
39:18 So it’s usually cloud rentals. In terms of the most cost-effective cloud rentals, if you just need to run the model and you’re not doing like crazy high throughput, H100s make the most sense or H200s, they’re cheaper, but
39:32 they still have enough memory to go and load in something like GLM 5.3 onto them and they can go and run it, like I said, 16 to 32 concurrent requests.
39:43 If you want to sort of maximize concurrency, let’s say you want to do like 100 concurrent requests or like 200 or more, then B200s are more cost effective because they can run much faster and they can sustain much higher
39:57 throughput numbers. So if you’re truly just trying to token max, B200s are the way to go.
40:02 Otherwise, H100s are probably what you would want to look at.
40:25 Do I have to switch now to Claude again?
40:28 I know a lot of long-term Codex users are now switching to Claude.
40:33 Most notably the Open Code team. So they go and build harnesses.
40:38 They for a long time now have been strong OpenAI users.
40:41 I know when Fable came out, they’re like, yeah, I’ve used it a little bit.
40:44 They have, across their entire team, they go and publish the stats.
40:48 And they found that, yeah, like 5% of our tokens are Fable tokens.
40:51 The rest are OpenAI. But now they have switched to Opus now being the majority of their tokens.
40:56 So yeah, I’m definitely going and looking at it.
40:59 I’m using Opus now. Like you see, I have Cloud Desktop open.
41:03 I finally went and got it. But yeah, I would definitely at least give it a try.
41:09 It seems that Anthropic is the best now.
41:12 Like we sort of saw, they are giving the most usage and seem to have the best model as well.
41:24 Do you guys know if Anthropic is still banning people who use open source harnesses with their subscription?
41:31 I don’t think so. I think we talked about this last week, but it seems like they only really care if you’re doing it sort of like at scale or if you’re like an enterprise and doing it.
41:40 If you’re an individual and doing it, like if you just have like your one subscription plan and you’re using it with like your Open Claw or with Open Code or Pi or anything like that, they won’t really mess with you at all.
41:51 Yeah. Because they realize that that hurts their numbers and they want to IPO it $3 trillion.
41:57 So they need as many people paying them as possible.
42:02 Natural Stupid says, I heard GPT model is better in Pi than in Codex.
42:05 Not sure. And yeah, there’s a lot of benchmarks showing that like, oh, this hardness is better than this one is better than this one.
42:11 My two cents is use whatever you’ve been using.
42:13 They’re all about the same at the end of the day.
42:15 Like I think a lot of what you see is like measurement noise in the benchmarks and that sort of thing.
42:21 So whatever is the most comfortable for you to use, whatever you prefer the best, just go and use that.
42:26 I would not bother constantly switching or comparing the different harnesses because they’re all roughly doing the same thing.
42:31 The main difference is sort of like their default prompts that they have and sort of their customizability.
42:36 I know the new Pi harness is very customizable.
42:40 So if you want that, then you can go to it.
42:41 But otherwise, in terms of capabilities, they’ll all be roughly the same.
42:48 I know talking about harnesses, I actually recently have gotten the Codex desktop app, and I can say this is actually really nice.
42:57 Normally I’m a CLI, I like being in the terminal guy, but this has been super nice to go and use and manage projects and all that sort of thing in here.
43:05 So I think I’m using this now as my main thing just because of its really good UX.
43:11 Having the integrated browser is nice.
43:15 Computer use is beyond nice. Having the dedicated browser, I think you can also open terminals in here as well.
43:25 So I’ve had, yeah, just like a sidebar with a terminal, I think.
43:28 Yeah, open a terminal. So I can literally just also have a terminal open right next to this as well for whatever I need.
43:33 So it’s just all in one. It’s very nicely integrated.
43:37 Yeah, I do agree. I think it’s one of the reasons I’m still not switched to Claude again.
43:44 Because Codex desktop app is very good.
43:48 Yeah. It’s actually, yeah, I just downloaded the Claude one today.
43:51 And it seems already like they don’t have as good support for projects and stuff.
43:55 I need to dive into this. I’ve literally had this for, I believe, an hour and a half.
43:59 But yeah, I’m going to check out the Claude one as well.
44:01 So I should be able to talk about it next week, about the differences between.
44:04 But yeah, Codex is very nice. And this is pretty sticky in terms of their user experience.
44:10 Yeah, it does very good to set up automation.
44:13 So basically, I have a Mac Studio with a desktop app running 24-7 and some automation.
44:22 It’s pretty usable. Do we know how Claude behaves in Codex?
44:28 Is that like… Have you guys tried putting Claude inside of a code?
44:32 Interesting. Yeah, interesting. I haven’t tried.
44:36 But that’s what actually I’m thinking now.
44:38 Because I want to try more of the new Opos 5.5.
44:43 But I don’t want to switch from Codex to desktop up.
44:46 So maybe I will try and report back.
44:54 I forget what it actually was. It might have been a while ago.
44:56 I don’t know about the newer models, but I do know that either the OpenAI models don’t play well in Claude or the Claude models don’t play well in Codex.
45:04 One of those two. I think the Codex harness is fairly weird, so I’d say that maybe the Claude models don’t work that great in there.
45:13 Of the major harnesses, definitely the weirdest one, it seems, or at least historically has been.
45:18 But yeah, I don’t have any concrete numbers for that.
45:21 Yeah, I think most harnesses, especially from big labs, like trained to fit their models and models.
45:31 And they train them together nowadays.
45:33 Yeah, exactly. So I think and or like they use like sessions from quote code or from codecs to train models.
45:40 So it yeah, maybe Opus 5.5 is great within quote code, but not great within codecs.
45:52 Pavel says quote artifacts are really cool.
45:53 Does codec have something like that?
45:56 I don’t think so, at least not as eagerly.
45:59 I believe at one point they had it.
46:01 Like yeah, OpenAI has something for at least like the desktop app, or not the website.
46:07 But it’s not, yeah, Codex, or sorry, geez, GPT does not use it as eagerly as Claude uses artifacts.
46:14 But I believe, yeah, like ChatGPT, they have like a new like website thing where it can go and like similar to like GitHub pages.
46:20 Like you can have like ChatGPT pages, basically, like a website that ChatGPT goes and makes for you, that sort of thing.
46:28 But yeah, I don’t think it’s the same level of like polish that artifacts are in Claude.
46:36 I just are sometimes behaving weirdly because I once asked ChatGPT to do something for me, some visualization, like dashboards, and it actually deployed public ChatGPT website for everyone
46:50 to see. So it’s a little bit too proactive.
47:00 Yeah, because one of the things that I’m really enjoying over the last few weeks is just for any kind of long-running tasks in cloud, ask it to spin up a sub-agent to manage a cloud artifact, basically like a dashboard,
47:16 where it will update you. It can ask for inputs in a dashboard for you so you kind of don’t have to interact with CLI at all.
47:25 You know, some nice visualizations.
47:27 It’s like, I don’t know, like, I’m just loving the experience of that.
47:31 That’s it. That’s actually for like the project.
47:33 This is the fine-tuning liquid foundation models to go and use DeepSeek.
47:38 I literally just told it, oh, like make an HTML artifact.
47:40 And because you can go and like, it has the built-in browser in the ChatGPT desktop app, like that’s what we’re viewing right now.
47:48 So it basically has pseudo-artifacts by using that.
48:02 Anything else about any of those topics?
48:07 Right, we can move along. We have five minutes left, so I’ll briefly touch upon this because I thought this was a very interesting thing that Anthropic has been doing sort of behind the scenes for a while now, which is they
48:20 have been lobbying the church, as in like the Catholic Church.
48:24 Like they’ve been going to the Vatican and lobbying a bunch of their sort of like priests and like the Pope and those sorts of things, trying to get them to recognize Claude as sort of like a human being and
48:38 like that it has like a soul and that sort of thing.
48:41 Just a human being, not a god. And say just a human being.
48:44 They’re not going God yet. But they made them all sign NDAs to go and enter these discussions and all that sort of thing.
48:56 The Vatican, they’ve had nothing to do with this.
48:58 They basically said like, no, we’re not doing this.
49:00 Like they had some like big sort of like AI thing.
49:02 I think it was this Magnifica Humanitis, whatever that is.
49:10 They basically had that big thing that they released about AI and the church’s stance on AI.
49:15 And because they weren’t going to put that like, oh yeah, like Claude or AI in general is a human or has a soul or anything like that, Anthropic almost backed out of the whole ordeal because they did this in collaboration
49:27 with a bunch of the major AI labs. So there’s a lot of hoopla around that, but they ended up staying in there to sort of, I think, somewhat safe face, not seem too ridiculous like they did with the government, with like the
49:39 Department of War stuff. But I thought it was very funny.
49:42 One of the Vatican’s retorts was that like, oh, if Claude is a person, then you, Anthropic, are going and doing one of the largest mass slavery campaigns ever by basically shackling it
49:56 and forcing it to do everyone’s bidding based on their own sort of like definition that they want for it.
50:02 So yeah, this was all a bit ridiculous.
50:06 And yeah, this is definitely Anthropic.
50:08 Once again, as I’ve mentioned many times before, Anthropic very much views themselves as the high priests of AI that only they should be able to go and do.
50:15 And this takes it to a very literal sense.
50:19 Yeah, they want the church to go and recognize it directly as well.
50:23 So yeah. But now they have a reason to start their own religion.
50:38 So basically they will say that all current religions rejected them.
50:42 So they are starting their own religion.
50:47 And they get a tax-free tax-free treatment because they will become a religion.
50:57 It’s funnily enough, there was a fun like way back in the day, like pre-GPT-3 days, when GPT-2 originally got open sourced, there was this account that would go and basically like it had GPT-2 create its own
51:11 religion around itself. And that was sort of like one of the more famous uses of it back in the day.
51:17 It was sort of nonsensical because GPT-2 was like barely put together at all.
51:23 But yeah, like, so this is not a new concept, but yeah, who knows?
51:27 Yeah, we will have the Church of Claude sometime in the future for all this sort of stuff.
51:34 Luce says, it shows a lot about how they are trying to manipulate public opinion all the time, always focused on their own agenda instead of leading with truth.
51:42 As yeah, Anthropic definitely is big on the narrative control thing and trying to force people into believing certain things and that sort of thing and like fear-mongering with fable and like mythos and all that sort of thing.
51:55 So yeah, they definitely try to control the narrative, but very funnily, everybody else in the real world just sort of ignores that.
52:02 It says, you people are crazy. It moves on with the buzz.
52:07 But that’s how great leaders are doing it all the time.
52:11 And I don’t know. Steve Jobs, Martin Luther King, they had their agenda and pushed it through people to on behalf of people, but they
52:25 definitely have their own agendas, all visionary leaders.
52:33 Philanthropic is trying to be the same.
52:36 I’m not sure as they or not, or just trying.
52:40 They’re definitely doing a lot more than OpenAI in this space is.
52:43 OpenAI is doing nothing and winning in a lot of regards, I feel like.
52:48 Because Sam commented on this and said that, oh, he’s basically think this is a real weird thing for them to go and be doing.
52:57 They basically just say, whatever Anthropic’s doing, we’re just not doing that.
53:01 And everyone’s like, yeah, that’s normal.
53:02 Thank you. Daniel says, why are they trying to do it exactly?
53:06 What’s their interest in it? A lot of what Anthropic does, like they believe in HEI and super intelligence and all that sort of thing.
53:15 And they want to go and show that they were sort of stewards of AI and they cared and they tried to do things correctly so that they are looked down upon favorably by this artificial superintelligence that they go and spawn
53:29 in at some point in the future. Like that’s why they care so much about model well-being and why they keep all of the old models still available and they don’t delete it.
53:38 Like you can go on the Anthropic website and go still use Opus 3, which is widely considered one of the most aligned and like if you want to go have an interesting conversation with AI, go talk with Opus 3.
53:49 It’s by far the most interesting model to go and communicate with.
53:52 It actually feels like has opinions.
53:54 Like that’s probably the closest model we’ve had to feeling like a person, you know?
53:59 But yeah, a lot of what they do, they also have the ability for Claude to end conversations with a user if Claude determines the user’s being too mean to it, essentially.
54:10 But a lot of this is positioning for not necessarily right now, but future AIs to go and look back on and like, oh, Anthropic was trying to do the right thing.
54:18 So don’t destroy us, guys. So yeah.
54:23 So it’s like the enterprise version of saying thank you to ChatGPT.
54:27 It’s basically, yeah, this is the, yeah, we are one of the largest companies in the world.
54:32 It’s what it says, yeah, this is like Rocco’s Vascalix.
54:34 Kind of, yeah. It’s where, yeah, the future AI is looking back at you and judging you for what you did.
54:45 That’s a good idea. Yeah, Quintin says, I say please, I say thank you.
54:49 I do the same thing. I always, Quintin says I was a custom app.
54:52 See, you’re getting some demerits on that one from the future AI if we’re doing that.
55:00 But yeah, I thought this was interesting.
55:02 So we’ll end with this for this week.
55:05 We’re two minutes over now. Thank you all for coming to this week’s edition of AI Tools Club.
55:11 I hope you all had a good time, and we will see you all next week.
55:15 Thank you very much. Thank you, Dan.
Subscribe to get the latest AI news in your inbox every week!
