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: Anthropic’s intelligence report catalogs more than 20 cases involving Claude, including hacking, social engineering, disinformation, biological threats, and weapons-related projects. The report explains how Anthropic detected and stopped some of this activity, but it also raises questions about the scope of its monitoring.
Takeaway: AI misuse reports can reveal real threats, but they should also be examined for what they imply about provider access to customer activity.
Links shown: Anthropic report announcement, Anthropic threat intelligence report
What was talked about: Alibaba, Moonshot, DeepSeek, and ZAI appear in claims about extracting Claude reasoning traces or capabilities for model training. The most serious allegation is that Moonshot and DeepSeek silently sent some user requests to Claude, which exposed sensitive Chinese and Russian information to an American provider while presenting the responses as outputs from Chinese models.
Takeaway: Silent routing is more than a model-quality issue because it can transfer sensitive data across providers and national jurisdictions without user consent.
Links shown: Anthropic threat intelligence report, NetAskari post on the Anthropic report
What was talked about: The detail in Anthropic’s report creates uncertainty about how activity was observed when API traffic was expected to have zero data retention. Possible explanations include safety classifiers, external monitoring, product-specific retention rules, or contractual loopholes, while speculation about training on customer data remains unverified.
Takeaway: Do not treat a zero-retention label as sufficient assurance; verify the exact product, exceptions, monitoring methods, deletion period, and training terms.
Links shown: Anthropic threat intelligence report, Anthropic report announcement, Tautologer post on DeepSeek and Moonshot routing
What was talked about: A likely reason for routing difficult requests to Claude is to obtain better answers and collect valuable reasoning traces for future training. A classifier could identify unfamiliar or difficult prompts, forward them to a frontier model, return the stronger answer to the user, and add the result to a continuously improving training set.
Takeaway: Model routing can improve both immediate output and future training data, but it must be transparent and must protect sensitive prompts.
Links shown: Tautologer post on DeepSeek and Moonshot routing, Anthropic threat intelligence report
What was talked about: Several methods for identifying hard requests are compared. A system can measure similarity between an incoming prompt and known training data before generation, inspect perplexity or uncertainty during generation, or analyze branching tokens and internal model behavior after generation. Fast similarity checks are especially suitable when routing must not add noticeable latency.
Takeaway: Use lightweight out-of-distribution checks for real-time routing and deeper uncertainty analysis for later evaluation and training.
Links shown: Anthropic threat intelligence report
What was talked about: Croft AI, a low-cost OpenRouter provider, was exposed for serving cheaper models from other providers instead of the requested models. Reports indicate that the problem remained after repeated notification, and the provider later disappeared from OpenRouter. The case also highlights broader concerns about budget providers retaining or reselling data despite zero-retention claims.
Takeaway: Choose inference providers by verified reliability and data practices, not price alone, because unusually cheap service can hide model substitution or secondary data monetization.
Links shown: Ryan RGS post about CrofAI, CrofAI is an OpenRouter wrapper
What was talked about: Model substitution can sometimes be detected through tokenizer differences. Model families often produce different input token counts for the same text, so billed token statistics can be compared with the expected tokenizer. The same technique has also helped identify unnamed stealth models.
Takeaway: Record token counts and compare them with reference tokenizers as a basic check that an inference provider is serving the claimed model.
Links shown: CrofAI is an OpenRouter wrapper
What was talked about: Darkbloom uses a network of Mac minis to serve local AI models and pays machine owners according to the tokens they provide. The service appears to support more models than before, but there was little current usage data or direct experience available to assess its economics and reliability.
Takeaway: Distributed consumer-hardware inference is promising, but evaluate demand, payouts, supported models, and operational risk before supplying hardware.
Links shown: CrofAI is an OpenRouter wrapper, Darkbloom network statistics
What was talked about: OpenAI reportedly moved 25 percent of its production engineers onto security work and used Astra to inspect production systems for vulnerabilities. Engineers then reviewed and fixed the reported issues. The approach reflects a growing need to harden systems before similar AI tools make vulnerabilities easier for attackers to find.
Takeaway: Allocate meaningful engineering capacity to security and use AI scanning as the start of a human-reviewed remediation process.
Links shown: a16z post quoting Greg Brockman
What was talked about: Codex, Claude, Daybreak Blue, Codex Security CLI, dedicated security products, and pull-request checks were considered for vulnerability discovery. These tools found inconsistent endpoint protection, overly broad permissions, and requests capable of taking down services. Identity verification can unlock defensive models, while normal models may refuse red-team requests because they appear malicious.
Takeaway: Combine multiple security tools with human review, clear system boundaries, least-privilege permissions, and automated checks on every code change.
Links shown: Expanding Daybreak as the Cyber Defense Window Narrows, OpenAI Codex Security CLI announcement, AI at Aikido, Claude Security plugin documentation
What was talked about: Gastown became popular as a large agent-swarm architecture, but its creator later described it as a failure for practical work outside building Gastown itself. The framework consumed many tokens, produced weak code, and was replaced in the creator’s workflow by a single capable agent.
Takeaway: Do not adopt a multi-agent framework because it is popular; test it against a single strong agent on real tasks, including cost and output quality.
Links shown: Dan Luu post about Gas Town, Dan Luu on AI-assisted coding
What was talked about: Newer models with effort controls can plan, implement, test, and verify work for hours without the external orchestration that earlier coding models required. The remaining orchestration gap is longer-term persistence across days or weeks, including memory, recurring maintenance, backlog selection, and management of token budgets.
Takeaway: Use native single-agent autonomy for bounded work and reserve custom orchestration for persistent goals, recurring operations, and resource management.
Links shown: Dan Luu post about Gas Town, Dan Luu on AI-assisted coding
What was talked about: Multi-agent systems remain useful when each agent has a narrow role, controlled context, and a clear evidence source. One example traces the purpose of code through commit history, repository documentation, and Slack in parallel before a synthesizer ranks the evidence. Loosely managed swarms instead multiply context loss and drift.
Takeaway: Parallel agents work best for separable research paths with explicit roles and synthesis criteria, not as an unrestricted default for implementation.
Links shown: Dan Luu post about Gas Town, Dan Luu on AI-assisted coding, Maritime
What was talked about: Reusable agent workflows are difficult to share because low-quality skills create noise, successful flows are often specific to a codebase or harness, and proper evaluation is expensive. The closing model news covered GPT Terra’s planned retirement, possible GPT-6 replacements, smaller models becoming brittle about instructions, and shrinking practical differences between frontier and mid-tier models on routine coding tasks.
Takeaway: Turn repeated actions into skills, validate them with explicit success criteria, and choose the least expensive model that remains reliable for the workflow.
Links shown: Codex Marketplace, Nate Berkopec on skill slop, SkillsBench, Cursor pstack plugin, MIT NANDA Hackathon, NANDA Town, Microsoft Foundry model retirement schedule, OpenAI API deprecations
0:06 All right, let’s get started. Welcome back to AI Tools Club, everybody.
0:10 This week, we are going to be talking about the Anthropic Intelligence Report, an open router scandal where one of the providers was exposed for not actually serving the models they said they were.
0:21 We’re going to be talking about OpenAI’s new security posture that they have.
0:26 We’ll also be discussing Gastown and how that actually sort of ended up being a flop.
0:33 And then we will finish off with any extra time with Pacing the Frontier.
0:37 And yeah, Pavel mentions, you want to talk about the flybrains?
0:40 Yeah, we can talk about the flybrain being open source.
0:44 But yeah, getting into things here is the Anthropic Intelligence Report.
0:49 So this is basically Anthropic went through and combed through a bunch of user data, which we’ll get to that being weird to begin with.
1:00 But they went through and highlighted a bunch of interesting sort of malicious use cases for clot that people were using it for and how they detected and stopped it.
1:11 So yeah, this is a huge report. I think they documented, I think, over like 20 different cases of how Claude was being used from hacking to social engineering campaigns and then also for bio stuff.
1:26 So I would post this in the chat for everybody to go and look at.
1:31 But yeah, there’s a ton here. Most of them are interesting and neat to read about.
1:37 Like, oh, people are using Claude to hack things.
1:39 I mean, this has been happening for years now.
1:42 But I think the most interesting things that they went and discussed in here was actually from China.
1:47 So Anthropic previously had talked a lot about Anthropic’s distillation attacks and how they’re extracting chain of thought from Claude to go and train their own models.
1:57 So yeah, they’ve discussed Alibaba here, which they’ve mentioned before.
2:00 Alibaba seems to be the largest distiller of the Claude models, at least that Anthropic knows of.
2:06 It’s yeah, over 151 million exchanges.
2:09 This is probably order of magnitude, like hundreds of billions of tokens, maybe even trillions of tokens from them.
2:16 But we already knew this. This wasn’t interesting.
2:18 I’ve mentioned before that like these are useful, but not sufficient for training a model.
2:24 What is interesting here and has actually caused some fallout in China is that they have noticed that both Moonshot, which is the company that builds Kimi, and also DeepSeek have
2:38 been silently rerouting requests from their own models over to Claude models.
2:43 And this has had a few major sort of instances where companies or governments like the People’s Republic of China are going and using these quote-unquote Chinese models.
2:56 And instead, their requests are being routed over to CLOD to go and collect more data for it as needed.
3:04 And so I think, yeah, this is, yeah, PRC, People’s Republic of China technology company.
3:09 They were using DeepSeek, and as they were using it, they were sending a bunch of internal sensitive information.
3:15 And DeepSeek was just sending that right over to Claude instead.
3:20 And Anthropic was able to go and harvest it.
3:22 And Anthropic, obviously, American company, they have control over this data.
3:27 Now, all of a sudden, they can give that to the U.S.
3:29 government. They also found this with Russia and also the police surveillance, the People’s Republic of China police were also using this.
3:39 And so, yeah, they’re sending over basically a bunch of information there.
3:42 And same with Moonshot as well. And so this is a big deal because no longer like the distilling, like Chinese government does not care that their companies are distilling from American models.
3:52 They probably encourage it, if nothing else.
3:55 And they also talk about in here how, I believe, yeah, this is for ZAI, like how they’re distilling, you know, cybersecurity capabilities to make them better hacking models.
4:04 They don’t care about any of that. This, though, they do care quite a bit about.
4:09 Because this is their own and also Russia’s own sort of data being leaked over to US companies.
4:15 And DeepSeek and Moonshot are not communicating this at all.
4:20 And so because of this, it looks like this is sort of in the rumor mill.
4:23 I haven’t been able to go and necessarily verify that.
4:27 But it looks like the CEO of Moonshot got arrested by the Chinese government.
4:35 And so who knows what’s going to happen there if he’s just going to, you know, it’s going to be a Jack Moss situation where he just sort of disappears and, you know, the Chinese government is just going to take over moonshot.
4:44 But yeah, this has had tangible fallout in China because of this.
4:50 So yeah, this is anthropic. They’ve been wanting to pressure China.
4:54 It really wanted them to sort of slow down or, you know, shine them in a bad light.
4:59 They’ve always been trying to make China the enemy, but it hasn’t really been able to do much, at least on the U.S.
5:04 side of things. So now with this report, they have gone and turned it, sort of turned China in on themselves.
5:11 And now they’re taking out their own companies.
5:14 So yeah, there is also the sort of question of how the heck did Anthropic get access to all of this data?
5:21 Because all of this, theoretically, should be happening through the Anthropic API.
5:26 And so that should have zero data retention, except for the Fable models.
5:30 But many of the models that they mention in here are stuff like Opus.
5:33 I believe Opus 4.6 and 4.8 get mentioned in here a lot.
5:39 And so there’s the additional question of, is Anthropic really doing zero data retention or are they holding on to all of that customer data themselves to go and use instead?
5:49 Like people on Twitter have been highlighting that once, I believe it was like Notion and also Figma, once those companies started signing large deals with Anthropic, all of a sudden the Anthropic models got way better at
6:03 front-end design and designing in general.
6:06 So yeah, there’s a lot of people saying that Anthropic is actually retaining and training on the data that you go and send to them as well.
6:14 And this sort of seems, based on, you know, their large list of all the different behaviors and activities that they’ve captured, seems to suggest that that is the case, where they’re holding on to all of this.
6:27 And so, yeah, Quintus says zero for several values of zero, or very close to zero.
6:32 Yeah, exactly. Only your bad chats, or only the really good ones, do they save.
6:36 All the rest they just throw away. But yeah, so I think this is undermining like a bit of both sides.
6:45 So obviously, like China, they’re being undermined.
6:47 Their labs sort of, they’re definitely going to have a lot more scrutiny now from the Chinese government about the actions that they’re taking.
6:54 But also, like we just mentioned, Anthropic now is sort of under the microscope a bit of what are they actually doing with our data and are they holding up to it?
7:03 So yeah, that’s all I got on this. Any questions or comments about this?
7:11 Well, I look forward to their explanation between zero and non-zero.
7:18 You know, how they describe how this works.
7:21 I think it’s going to be one of those things.
7:23 They’re not going to mention it. Like, I combed through the report.
7:26 I obviously didn’t read everything because it’s like 100 plus pages effectively.
7:31 But yeah, I didn’t see anything in there of how they actually were capturing all this information, whether it was from classifiers going off and safety guardrails being triggered, or if they have some external monitoring.
7:44 My guess is they’re not going to release that or talk about that.
7:46 Like I said, they’re just sort of trying to, they’re much more trying to undermine China and highlight negative use cases.
7:52 And I think in a lot of ways, trying to highlight why they need to go and save this data to be able to go and catch this misuse.
7:58 Like, I do think there is a little bit of sort of a reason for them to go and do this.
8:03 And, you know, I think right now they have a 30-day retention for all fable traces, where they will save everything.
8:10 And then after 30 days, they will go and delete it.
8:11 And I think for a lot of safety concerns, like they highlight a bunch of stuff of, you know, like people going and doing mass information or disinformation campaigns to influence elections.
8:22 I think they talked about it happening across a couple of different countries on both sides of the political spectrum.
8:28 And then also, yeah, trying to build bioweapons and other weapons-based systems.
8:34 Yeah, we’re on like a whole section right here of trying to build autonomous drone swarms and stuff like that.
8:38 So stuff like that, I think they should be monitoring.
8:42 But yeah, the question is, how far does it go?
8:44 And are you also then going and training on this information as well?
8:47 And how do we go and verify that all of that is happening?
8:49 Because I think the AI field, we’ve shown, you know, most companies, both US and China, are going and pirating huge amounts of information.
8:58 And so I don’t think at this point we can really trust their words saying like, oh no, everything we have is above board.
9:05 So yeah. Brandon says, does even them catching China then mean that it goes against their terms of service?
9:13 And yeah, like I said, we don’t know.
9:16 They don’t mention it. And it’s been very weird that people haven’t been talking more about it, I feel like, at least that I’ve been seeing on Twitter.
9:23 But yeah. My guess is that maybe clawed code, they save stuff, because that’s technically not through the API.
9:31 So potentially all your clawed code chats and stuff can get saved.
9:38 But yeah, that’s yeah. My guess, yeah, they probably have at least a loophole within their terms of service would be my guess at the very least, if not outright going against it.
9:53 What is the like, I don’t know, reason, the rationale for them to be routing requests to Claude?
10:00 Because it feels a bit, you know, wasteful.
10:03 And first of all, like, would the, you know, the CEO of Moonshot not foresee that it’s not something that Chinese government would be happy about, who like are probably their biggest customer or something?
10:13 Like, I don’t know, what’s to pretend their model better than it is?
10:16 Like, what would it, why would they even do it?
10:19 It’s to go and capture better reasoning traces.
10:23 So a part of these distillation attacks is I’m also getting the reasoning for it to go and train models on.
10:28 So the sort of like what I would guess that they were doing, like I said, we don’t really have information, like the Chinese companies obviously haven’t talked about this.
10:37 But my guess is that they have like a classifier model that is going into detecting what would be a hard query or what would be a query that their models go and struggle with.
10:45 And then if it’s a hard question, they send it to a frontier model either from OpenAI or Anthropic to go and answer it instead.
10:53 And then they go and give that to the user.
10:54 The user is happy because they get a better answer than they would have potentially gotten from the Chinese model.
10:59 And then the Chinese lab, they now have a new reasoning trace to go and train on for their models to make it better in the future.
11:08 Because a lot of what these labs are doing, both in the West and also in China, is sort of a data flywheel, where they will go and find what are the hardest questions that were asked of the model, where did the model sort
11:21 of struggle, and then picking those out and then going and training on those and making better answers for them, that sort of thing.
11:29 So yeah, that’s why they were doing it.
11:32 My guess is the Moonshot CEO didn’t necessarily say to do this.
11:35 My guess is one of the engineering teams probably had a meeting where they’re like, oh, what’s the best way to go and do this, to go and harvest more data?
11:41 Like these labs all have data teams, and they’re always desperate to go and find new ways of generating, you know, unique questions and new data and real-world problems.
11:51 And this is probably one of the ways they went and figured that out.
11:55 So yeah, I think up to this point, most of the Chinese labs have been acting somewhat independent of the Chinese government.
12:02 Like the Chinese government really hasn’t sort of, you know, put their hands too deep into any of the labs, at least from, you know, what I’ve been able to see through investments or sort of like advising or telling them
12:13 what to do. They give very general guidance, I think.
12:15 The only thing I’ve really seen them say is that they need to open source their models and that like, oh, we will be a pro-open source country, essentially.
12:24 But beyond that, I have not seen too much from the Chinese government.
12:28 So I don’t think they think too much about their involvement with it.
12:31 But I think from now on, that will be a much bigger factor in how they operate.
12:37 So why would they like risk this situation?
12:43 Because they can still do the same thing as they did before, and they just not show the output of Claude to the user, just kind of fail the user in the real time and then later save this and get it to Claude.
12:55 Is it just like, you know, because they would get that reasoning trace either way.
12:59 They just, the difference is like show it to the user or just keep it internally for distillation.
13:04 Is there a good reason to show it to the user just to make the model looks better?
13:09 It makes the model look better. User is probably happier because they’re using a better model.
13:13 Like at that point, it doesn’t matter if you show it to the user or not or you save it.
13:16 The main issue with this is that the request is being passed back into Claude and the request has a bunch of sensitive data that is being sent through.
13:26 And maybe I assume it’s the same logic as what Claude Cod sometimes does when it’s asking you, can we use this session for our training purposes?
13:38 So when I talk to somebody in San Francisco about it, basically the model is measuring the perplexity of that session.
13:52 So basically how surprised was the model on average by kind of token that the user has submitted.
14:01 And I assume that’s what the Chinese also did.
14:05 So when the model got kind of very surprised, so basically this prompt was out of distribution basically for what the model is trained for.
14:15 They would kind of try to enrich the distribution by going to cloud.
14:21 I think it’s kind of a very natural thing to do because you don’t know what you don’t know, right?
14:25 You don’t know what kind of trajectory of a conversation would surprise your model before you actually encounter it with your users.
14:34 So I think that was kind of the natural thing to do.
14:39 And yeah, I think what you described there, it’s actually two different things you just described.
14:44 So the first one is measuring model uncertainty, which is, yeah, all these labs do.
14:48 Like when I say, you know, like Anthropic Achatipe, they’re looking for hard queries.
14:52 They’re basically measuring model uncertainty throughout the question.
14:55 And then when uncertainty is high for a given question or, you know, after a given sequence of tokens, they go and save that and say, like, oh, let’s go and like dig into this more train here potentially.
15:05 My guess is that because both DeepSeek and Moonshot, it seems like they are doing it before their model even starts generating any tokens, they are going and sending the request over to Claude.
15:16 And so to detect that, because they have access to all the data that they train on, they can go and measure the similarity of that question to all of the rest of the data that they have gone and gathered.
15:26 And if that question is not close or similar to any of the data that they currently have, then they go and send it to a plot or OpenAI model because they know that that question is out of distribution
15:40 for their own model. Yeah, but is it kind of the same thing, just different steps?
15:45 Because what I assume they did, they would have a classifier that would predict what trajectory is going to be most perplexing for the model, and then they’re just going to route it.
15:56 So rather than actually measuring perplexity, they kind of predicted perplexity in a way.
16:02 So that’s why I mentioned that. Isn’t it kind of the same thing?
16:05 It’s yeah, well, because I mean, your predictor, like I said, your predictor isn’t actually trying to estimate perplexity.
16:12 It’s yeah, it’s just measuring similarity to data that you already have.
16:16 Whereas, you know, if once the model is done answering the question, a lot of the times you won’t even use perplexity a lot of the time.
16:23 You’ll look at the output distribution and see, you know, like you’ll find tokens that are high ambiguity.
16:29 They’re called branching tokens. Where whenever these sort of like, oh, at this step, it could have gone like five different directions.
16:37 So it’s sort of unclear. That’s sort of like the basic thing that a lot of them will do.
16:42 And then they’ll sample more. But they also have a lot more advanced tools as well to go and look at it after the fact where they actually go and look and do some sort of like analysis of how the model is thinking.
16:51 And so they do a lot more like mechanistic interpretability of the model to measure it as well.
16:57 Hmm, super interesting. Cool. Kind of funny that Anthropic, you know, for some requests downgrades you from Fable to a downer model, and Moonshot would upgrade you from Kimi to
17:11 a smarter model, or at least what they think is a smarter model.
17:16 Yeah, exactly. Anything else about this?
17:28 Any idea how exactly they identify what are the models that need to be brought to me?
17:42 Because I feel like the mechanism for doing that has got to be pretty smart, right?
17:46 It’s, I mean, you want it to be relatively fast and low compute.
17:49 Like I said, like the way I would do it, like naively, just thinking of it on the fly, is you just measure the incoming question to the set of training data that you have across all your RL environments, supervised fine-tuning,
18:02 and pre-training. And you basically have a threshold where, oh, if this question is not within a certain similarity level of any of that data, then you assume that your model has not been trained on that sort of question,
18:13 or it would be bad at answering that sort of question, and then send it over to a Quad or OpenAI model.
18:19 Because for those types of models, you want them to be quick, because otherwise it’s going to impact the user experience fairly quickly in a negative way.
18:32 I feel like honestly, doing that, and OpenAI should just do that more.
18:39 Maybe route the same element itself.
18:42 Yeah. Sadly, because they’re the frontier models, who do they route to?
18:46 There’s nobody left. But yeah. All right.
18:53 Moving on. Sticking with scandals right now and models not being the models that they say they are.
19:01 This is something that happened yesterday, actually.
19:04 So there was this provider on OpenRouter called Croft AI.
19:10 And so they were actually one of the most cost-effective options on there.
19:15 But it recently came out that they were not actually serving the models that they said they were serving.
19:21 And instead, they were routing you to cheaper models hosted by other providers instead.
19:27 And they basically, they got caught.
19:29 Their founder blew up on Twitter. You can see they went and deleted their account.
19:33 They’re no longer on open router or anything like this.
19:36 But they had been, you know, sort of quote unquote reliable and like a decent number of people used them and relied upon them on open router.
19:44 So yeah, this person went through. I’ll post this in the chat.
19:48 But yeah, they went through and basically found out that what models, when you’re requesting a model from CroftAI, they were sending them to a whole bunch of different other ones.
19:58 And they were able to go and detect it all.
20:01 And then they also reached out to the team at CroftAI, and they basically were like, huh, that’s really weird, and never fixed it, even though they were given multiple attempts to go and fix it.
20:12 So they have all the chats and history here of that.
20:18 So yeah, they place Blue Up, but this is sort of a warning.
20:22 And you need to be aware if you’re using stuff like OpenRouter.
20:27 Make sure that you can trust the provider that you’re using.
20:30 A lot of the times, if you go and see a cheap provider, something is probably off.
20:35 I know this was something last week, actually, that the Open Code team also discussed, where a lot of the providers they were talking to, they do a lot of sketchy things.
20:46 Like they will say, oh, we have zero data retention policies, but in actuality, they go and save all of your information and then are reselling it to other AI labs to go and use for training as well.
20:58 And this is like fairly common practice.
21:00 And from what it sounds like, more often than not, that that is the case, especially when you have these sort of like budget sort of inference providers.
21:09 Because right now, inference for these models, especially these very efficient, you know, like DeepSeek V4 flash and GLM 5.3 flash, there’s a high demand for them, but there’s also a huge supply of people serving them.
21:22 And so the prices and margins are razor thin.
21:25 Most of them are probably actually net negative in terms of profit when running these models.
21:32 So they have to go and recapture that value elsewhere.
21:35 So yeah, if you’re using this Molten Open Router, just be aware of this.
21:39 Go and, you know, if you can try and measure, you know, if a model is working or not, or you know, like at least sanity check it, that it’s within the bounds of what you would expect it to be, I would highly recommend doing
21:51 that. But yeah. Has anybody ran into, does anyone know if they actually use Prof AI on open router at all?
21:59 Oh yeah, their website’s also bad. Or run into any issues with OpenRouter in general?
22:09 How do people catch it? I believe that there’s a couple ways you can catch it.
22:15 The first one that meets, I didn’t actually read through the end bit.
22:18 I believe they go over how they actually catch it.
22:21 But one of the ways you can go and do so is by checking the tokenization.
22:26 So all the different families of models will use different tokenizers.
22:30 And thus for the same sequence, you will often get different token counts that will be off slightly.
22:37 So for instance, if you’re using, let’s say, DeepSeek V4 Flash for a given input, you would expect it to go and use like 300 tokens.
22:45 But if you were sending it to a GLM model, it would actually be billed at like 294 tokens instead.
22:53 And so you’re able to go and line it up based on input token counts based on that, because sites like OpenRouter, and I believe most of the other providers as well, they will give you very granular statistics about how you’re
23:04 being charged. Like how many input tokens were you build for requests and how many output tokens.
23:09 So that’s like the most basic way. Like whenever you go and see a stealth model out there, like GLM 5.3 Flash was a stealth model for a while, but people were able to figure out within a day that it was probably a GLM model
23:22 because the tokenizers were the exact same.
23:25 So that’s the way you can go and do that.
23:29 I don’t know if that’s what they used here, but that’s one of the ways.
23:37 Pavel says, speaking of OpenRouter, did anyone actually use Darkbloom since it went viral?
23:43 I have not personally. I don’t have a Mac Mini that can go on there.
23:46 For those that don’t remember, Darkbloom is a way, if you have a Mac Mini, they are an inference provider that’s basically a network of Mac minis, all running inference for a bunch of local AI models.
24:00 And then they go and pay you out based on the number of tokens that you serve.
24:05 So if your Mac Mini can serve a lot of tokens, if you have a decently beefy Mac Mini, then you can get paid a decent amount.
24:13 So yeah, as we talked about this for a bit, it seems interesting, but I’ve not followed up with them at all.
24:21 I think a few people in that meeting said that they were going to try it.
24:25 But yeah, it looks like it supports more models, which is cool to see.
24:36 It doesn’t seem like the data is there.
24:40 All right. If there’s nothing else to talk about that, we can move on to the next topic, which I thought was very interesting.
24:48 This is from OpenAI. So this is one of their podcasts they did with A16Z, big venture capital firm.
24:55 They basically said they took 25% of their production engineers and put them on security and upgrading their security using Astra.
25:04 Just basically going and using Astra to go and diagnose all of their production systems, come back with all the bugs, and then their job was to go and filter through and go and actually fix all of them.
25:14 I thought this was very interesting, like to see how these large companies are going and handling this.
25:21 And also gives a good metric for us, I feel like, as we’re building different applications and trying to deploy it, how much we should be paying attention to security as well.
25:30 Like I’ve said this over to a lot of my enterprise clients that I go and talk with and basically said like we probably need to start beefing up the infrastructure.
25:39 Like if OpenAI is, you know, they’re the ones making these models.
25:43 So the reason they’re doing this is because, well, their models can go and detect these issues.
25:48 So if they’re not patched, you know, a third-party version of Astra could potentially come through and find these issues as well and exploit them instead.
25:58 So yeah, I wanted to hear from the community here if anybody is actually looking at the security of their code at all that they go and deploy.
26:08 And if so, how they’re tackling that and using that.
26:12 I’ve used the codex blue level, which found a bunch of things in my code.
26:21 Nothing on fire. And prior to that, I had pointed, I forget what version of Opus or a Cloud model at my code, which found a lot of inconsistencies
26:35 in how endpoints were secured. I would say that’s the common pattern, which sounds like Vibe-coded slot, of course.
26:45 But it found those inconsistencies, and then I ran the same test.
26:50 I found it very valuable, and it’s really my only option for security since I’m a solo working on this stuff.
27:01 I know for, I did it probably about like three, four months ago for a compute community, where I did a few iterations of sort of like adversarial code review there.
27:11 And yeah, I kept finding a bunch of sort of like, some of them were like mild issues.
27:15 Some of them were like, oh yeah, it can just bring down the entire service.
27:18 You know, someone could send a malicious request that, if formatted correctly, can bring the whole thing down and stuff like that.
27:24 So it’s definitely, from my experience.
27:26 It might bring down civilization. That’s right.
27:28 It’s yeah. But yeah, it seems to be very useful.
27:32 But yeah, Brandon, it seems like you have something to say.
27:35 Yeah, a couple things, actually. One, have any of you guys run into any issues with like, I know, for example, I tried to do a little bit of red teaming myself.
27:45 And my code actually hit like blocks and it like it was kind of like oh like are you trying to do use the coders for malicious purposes?
27:56 Like I need to like apply for the actual like higher gear.
28:02 So I kind of just decided, you know, I’ll reboot this later.
28:09 Brendan, did you use Daybreak Blue?
28:13 Oh, no, I don’t think so. But what is that?
28:17 So this is some sort of model that is set up.
28:19 You have to go through some registration process.
28:22 As a sole, as a single person, it was easy to go through.
28:26 An enterprise has a much higher bar, which I would never go through.
28:32 But I did not run into that accusation of hacking any other system because I guess the way I registered the model knew that that’s why it was registered to me, was to look at this code.
28:49 Was that a clear explanation? I don’t know if it was.
28:58 But yeah, it was telling me not to.
29:01 I was just thinking, you know, like, I would worry about that later.
29:06 We actually, funny enough, I think there are like a bunch of people who, either from like people in our company, like cohort who are trying out the product or like hopefuls, like students who are
29:20 looking for jobs. Like they were kind of just doing some type of testing for us, just like on their own, kind of like get our attention.
29:27 And so they just would post in Discord saying, hey, we get by this vulnerability.
29:31 And we would just want to make some mistakes.
29:36 So that helps. But I think that most of the security vulnerabilities I’ve encountered, at least in maritime, in general, are kind of just more so like
29:51 systems that… There’s kind of a lack of clarity, right?
29:56 We were optimizing for speed, and so we didn’t really check exactly what was doing certain things, right?
30:01 Or what claimed code was doing certain things.
30:05 And so I feel like the best kind of insurance against security is more like, how do you orchest the system while you have clarity on what exactly the system is supposed to be doing and how it’s doing it so
30:21 that you’re not really tearing in and keeping openings because of band-aid fixes?
30:27 Right? Like, and a lot of times it’s just kind of like fixes that allow certain things to go in, but like you’re a little bit too broad on the permissions.
30:38 It’s kind of like you allow like certain certain of your own services to access certain networks.
30:44 But like you’re a little too permissive with the with the way you allow those things to access.
30:49 So then it allowed people, just random people online to also use that.
30:55 So yeah, I would say like ultimately in my experience, the security vulnerabilities are more about kind of like human error than like anything, like any issues with the specific
31:10 setup of the software itself. It’s yeah, I do know.
31:16 So yeah, going to the initial question of yeah, yeah, OpenAI Daybreak Blue specifically, I believe is the defensive one.
31:24 It’s yeah, you can sign up for that, I believe.
31:27 You have to verify your account, which is basically just you need to provide a driver’s license or a passport or something like that, and just associate that with your account and you will be able to have access to the model.
31:37 And then there’s also the Codex Security CLI, which isn’t a chat interface, but it will go like statically sort of like it runs a default prompt through their model to go and analyze everything and come back with different
31:50 vulnerabilities that they go and find in there.
31:53 So those are two different tools that you can go and use to catch vulnerabilities.
31:58 Anthroopic, to my knowledge, does not have any tool that would allow you to do something similar to this.
32:05 I see. So it’s kind of just more so that they just want to make sure that they, first off, that you’re a U.S.
32:11 citizen, and two, that if you are using justice purposes, they have someone kind of point out.
32:17 Yeah, I don’t even know if you necessarily need to be a U.S.
32:20 citizen. I just think, yeah, they want an actual identity in person to go and associate if there’s any issues with the model, you know, like somebody to go after if you’re using it maliciously is mostly what they’re looking
32:33 for. And Pavel also says there’s dedicated AI solutions as well.
32:38 He’s using a Kaido. It’s yeah, I assume, yeah, there’s probably like dozens of them.
32:44 I assume like every YC batch has somebody doing it.
32:54 But yeah, I was going to say also the I know I’ve in the past I’ve gotten away from it, but I used to have automatic checks.
33:04 So whenever I would go and open up a pull request on GitHub, for instance, it would go and automatically do sort of like a scan for bugs and also for security issues as well.
33:14 So that way, anytime I’m like making or shipping new code, I have some process that goes and analyzes it for me.
33:22 So I know that’s an easy way to do it in instead of having to go and use an external tool like this as well, if you’re interested in that.
33:30 I think Quentin, I believe, has talked at length about that in previous AI Tools Clubs episodes as well.
33:39 And Brendan, there’s a cloud security plug-in that it actually seemed to run almost as long as the Daybreak Blue scan from Codex, from OpenAI.
33:51 I have no way of judging one versus the other, but it certainly spent a lot of time ranking through my code.
33:58 So there may be something that you could use for Manthropic as well.
34:03 Good to know. I’ve not heard of this before.
34:07 Yeah, I mean, I think actually, if there’s anything I’ve learned from example of the way PSAC handles stuff, but in cursor, probably better to just literally have the double check for both models.
34:21 As long as you’re okay with trusting Claude and Codex for I mean, no one knows how they caught China, but they did somehow, so clearly not zero-day retention.
34:37 Yeah, anybody else have anything else they want to comment about for AI security stuff and how they’re using AI?
34:50 If not, we can move on to which one is it?
34:56 This. So I wanted to do sort of a post-mortem on this because we discuss a bunch of tools and whether or not they’re good here at AI Tools Club.
35:05 But I feel like we rarely go back and actually sort of assess whether or not they were good or not.
35:11 So I thought this was very interesting.
35:14 So Gastown, this was big news, I believe it was like the start of this year, if I remember correctly, where it was this big agentic swarm sort of architecture that would, I called it a token torcher, where it just used a
35:27 bunch of tokens, wrote a bunch of slop codes, sort of did a bunch of everything.
35:31 But I felt didn’t work very well. I believe we had a few people sort of go test it out, but it was viral.
35:37 Everyone was talking about it on Twitter, that sort of thing.
35:40 But in a recent blog post, the creator of it, Steve Viegi, he basically said that the product itself, for his own usage, was a flop.
35:51 Like the only thing that Gastown was used for to go and build was itself, and that trying to use it on any other project didn’t work, and it sort of fell apart using it elsewhere.
36:02 And that he has since moved on from it because he had this big sort of like framework.
36:06 He had previously made beads, which he was like, oh, that was like a precursor, you know, a lesser version of Gastown.
36:11 And he envisioned, you know, these future sort of like more ambitious, even more autonomous, more token slop versions of Gastown.
36:21 But it appears that he moved away from them entirely and now he’s just using, you know, a single agent.
36:26 Like he talks about he just uses Fable 5 and just uses that to go and build some game that he had made in the past and is working more on that.
36:36 So yeah, I thought that was very interesting to go and see because this for a while was the big thing of like these agent swarms and like, you know, spawning in all these sub-agents and everyone having a different role and
36:46 doing all these different things. But I think as time has gone on, people have found that using only a single agent and having that go and sort of do everything is actually faster and provides higher quality code
37:00 than these larger scale solutions have.
37:03 So I thought that was very interesting to go and see sort of like the post-mortem from the creator himself about this.
37:10 But with that, I wanted to see, is anybody else using sort of like these big multi-agents forms?
37:15 I believe Daniel, you had talked about like you’ve built your own with these.
37:18 And like, do you sort of like feel the same thing as Steve?
37:21 Or have you been able to get stuff out of these multi-agents forms?
37:26 Yeah, actually, I noticed like totally similar vibe, basically, but also because of the subscriptions squeeze.
37:35 Like I totally feel the subscriptions.
37:37 Like, you know, half a year ago, I did not feel my codec subscription draining.
37:42 I could like, you know, have a few parallel agents.
37:45 You know, I didn’t care that it was like using 10x tokens.
37:48 I only cared that it didn’t bother me for like 20 hours and it was actually done because it just kept looping on itself.
37:54 But now it’s like really easy. Like I accidentally today use 70% of my codec subscription just like that on some like background thing.
38:03 And also one thing I noticed that impacts this is that, especially with GPT models, is that they’re now trained with this like effort parameter as of it wasn’t the case actually when GasDown came about.
38:16 There wasn’t effort in January. It was just the agent would do something and they would be like, hey, I did this and something.
38:25 And when they added the effort, and when you run something now on like any GPT model at like high effort, and it’s like a convention now at this point, they just do all the things that their orchestration was intended to
38:36 kind of handle of like do the plan, do the like the stabs, do the prep, do the verification, do the testing.
38:42 And then they just like inherently already, an individual model can already run for like hours and hours and hours and knowing what to do.
38:50 And like they kind of don’t need that much orchestration.
38:53 So now I’m like just shifting to how do I like I’m basically using my hardness more to like orchestrate them across days so they like can retain their goals across days on like, you know, which can be, you know,
39:08 watch things for an Facebook marketplace for me to buy or something, which could be some silly things like that.
39:13 But that’s kind of where they haven’t caught up to because even like goals or like schedules, they are not very good at like maintaining persistent goal across like weeks.
39:25 But whatever this like gas down and orchestration like frameworks we’re trying to solve is like having them solidly work for a few hours and achieve the goal.
39:36 And now there’s the goal command and it’s kind of doing everything out of like natively.
39:43 So yeah, I feel like it was a good product for the time.
39:46 It was solving the problem and that was what I was doing as well.
39:49 But that was a different challenge.
39:52 Yeah, I think they internalized everything into the models themselves, seems like.
39:57 And so what it seems like you’re saying right now, is the issue like memory and long-term memory, or is the issue sort of like ambition and like the model gives up after, you know, like a couple weeks of like checking Facebook
40:10 Marketplace on that sort of thing? No, actually checking Facebook Marketplace is not a good example because there it like you can literally just set a you know a clawed or codex workflow where it like literally has a you
40:23 know rawn job or something to do it.
40:25 Something along the lines of like I have some scripts that I made the agents do for me and then they also maintain it and they also like implement features when I like report a bug or something.
40:37 Things like this that I try to like that’s not something you can out of out of the box get from Codex.
40:43 It’s not going to like weekly check in on like user reports, merge the PRs, run the tests.
40:51 It will do one thing at a time, but it won’t like do it weekly.
40:55 So they’re not there yet. So that’s where I’m like using my like building on top of this.
41:00 So that’s kind of the distinction. Before I also needed them to, you know, because I think from like 2025, AI usage for coding was more like very interactive.
41:12 And now it’s a lot more of like implemented this thing and it goes and it disappears for like a few hours.
41:19 So to use the 2025 models, you had to like orchestrate them by force and now they get orchestrated naturally.
41:28 And I feel like the next thing would be like these more persistent memories where they can like track their progress across weeks, make like decisions, manage their budget of tokens or percentage
41:42 points of the usage or something. That’s what I’m expecting them to be using because it’s pretty convenient to like, you know, I have this piece of software I built for myself and now I have that I like delegate 10 percentage
41:55 points of my weekly codex usage to maintain it.
41:58 And then the agents are free to see that they have unused budgets and they can go to backlog and pull something.
42:05 So that’s something that I’m trying to pull out of now.
42:09 That’s where like orchestration on top of the existing harnesses works, but I feel like they’re just all moving in the direction anyways.
42:21 So there is there’s yeah, so does that answer the question?
42:28 Nice, yeah, that does. Thank you. That’s yeah.
42:32 Any other thoughts on harnesses and sort of like multi-agent branding?
42:37 I actually have a very different view of it.
42:39 I actually think that if anything, agents were more relevant.
42:43 It’s actually something I use in my own workflows a lot.
42:48 And I would say like the caveat here is like if you if you’re having the agent manage its own orchestration, right, like and you don’t really have any, like and you’re having it also like provide the rules
43:02 for itself to orchestrate with and split and split up the agents, then it will deviate faster than if you just have one session, right?
43:11 Because at least that one session has the context, right?
43:14 But like in the case of like having the agent manage its own orchestration, it’s just forgetting, it’s just speeding up the rate at which it forgets the context, right?
43:20 And the rate at which it veers off course.
43:23 But at least like in my case, right, like for example, one of my favorite workflows that I actually have now is the workflow where I’m trying to just understand.
43:34 In the case of Maritime, we have 50 to 100,000 lines of code throughout the code base.
43:40 And it’s very difficult to keep track of why a certain line of code exists in the code base or what exactly it’s doing.
43:49 And so one of the skills I use is a skill that literally I just say, okay, flash Y does this line of coding.
43:58 And then it will spawn several sub-agents.
44:01 It’ll spawn an orchestra, like a synthesizer that then spawns a sub-agent that goes down different threads.
44:07 One will go through the GitHub commit history.
44:10 One will go through the repos documentation.
44:14 And another will go through the Slack channels.
44:19 And then each of those gathers evidence in its own way.
44:22 And then it synthesizes and then like ranks based on relevant.
44:27 And so I think like, for example, like in that case, right, it’s like, if you just had a normal, like in the case of just like the normal orchestrator flows, like what would happen, what tends to happen is the model only
44:38 goes like so deep and then just says, oh, found the smoking gun.
44:42 But there’s a lot of decisions that are baked in, right?
44:45 And a lot of times they’re very conflicting made historically.
44:50 And so I think that in general, swarms are kind of for if you have a pretty good idea of what each individual within the swarm is actually doing.
45:03 I find them as extreme. You’ve got it there, Brandon, by the way.
45:11 Oh, which part? It’s near the end, like your most recent sentence, I think.
45:18 Oh, yeah, yeah. Because I’m just saying, like, in my experience, just if there’s a clear idea of what each agent is specifically within this one, and like each, and there’s a very, very clear constraint around what its
45:32 role is and what context it has. I found that to be extremely effective.
45:39 Yeah, I agree with that actually, because that’s something that’s missing from all the harnesses, like codecs and stuff.
45:45 You can’t really design your own workflow from scratch.
45:49 You kind of can. You can do a skill that tells the agent to make a sub-agent and tell them to follow this skill.
45:57 And then, because in Codex, I think you can also increase the depth of sub-agents.
46:00 So they’re not like in Claude. So you can actually have a few graphs like that, but it’s a little bulky.
46:07 But that’s totally something where you could still have a very nuanced workflow of find where this line of code is.
46:16 And that’s one of boxed workflow, be a recursive scale or something.
46:22 And then in a different workflow, when you’re doing a code review, you’re like, for each high-importance line of code that was changed recently, run this workflow, and then you have another sub-agent for each of the lines
46:33 of code. It does a bunch of this. So yeah, it would be, yeah.
46:38 So that’s what I’m also experimenting with.
46:40 But I feel like it’s, I’m kind of surprised that it’s not as like common.
46:44 I wonder why to like share these workflows between people or like, because if it was as easy to build in Kapika, because I feel like everyone would be sharing their different like, as like packages of skills, not just an
46:58 individual skill, but like a bunch of skills that are connected and like to like interact with each other, sub sub-agents, sub-sub-agents, et cetera, rather than just like one skill.
47:09 There’s actually several reasons why this is challenging, right?
47:14 Like for one, I think that there’s just too much slop, right?
47:17 Like specifically skill slop. I think Andrew…
47:20 Andrew, were you the one who said you liked that term?
47:23 Oh, no, no, someone else. Yeah, but there’s just so much skill slop, right?
47:27 Like, because it’s so easy for you to just, for people to just crap out, like, like a vibe-coded, like, oh, do this workflow, right?
47:36 And then talk about how awesome they are.
47:38 Like, unfortunately, even Gary Taylor does this, right?
47:43 And so, but there’s no vetting, right?
47:46 Like, there’s no, it’s not easy to say, like, okay, like, here’s like the success criteria.
47:50 Like, this is why this skill is so effective, right?
47:53 And so, like, the only way for that to really work is for a company that has shown very, very clear results with workflow to actually host that, right?
48:03 And so, like, I think the big issue here isn’t really necessarily that companies aren’t doing this or that people aren’t doing this.
48:10 It’s just that there’s so much noise out there that it’s being drowned out.
48:15 And two, actually, like, this was kind of what I was mentioning last weekend with PSTack.
48:20 Like, I think PSTAC is actually a really, really good example of how Cursor and Grok.
48:26 And I mean, like, to this point, like, very much is a clearly meant to shill cursor.
48:35 But the why skill that I mentioned that kind of interviews and then figures out what’s going on, that’s actually something that I took from PSTAC and made my own.
48:50 And I do think it is pretty smart how, and I think I also saw something like a workflow that I want to say like CrowdStrike uses as well, maybe Cloudflare.
49:03 Like, there’s some guy who made a post about it.
49:06 Like, I do see there are some very, very smart flows that are arising out.
49:12 But it’s also kind of like, we’re getting to the point where it’s very, very bespoke workflows, right?
49:18 Like, they’re very, very specific to harnesses, and they’re very, very specific to the way, like kind of design philosophies, right?
49:26 And information of the actual code base.
49:30 And so I feel like it’s kind of more so like one of those things where I think that everyone’s going to eventually have their own set of workflows.
49:42 But it’s kind of hard to explain specifically how to have a fully working orchestrator flow.
49:51 But in my case, a lot of it is just me just realizing that I’m doing action and asking the agent to do certain things a lot.
49:59 And then just converting that into a skill.
50:02 And then saying, okay, the next time you detect that I’m trying to do this sequence of action, just go into that skill.
50:13 I hope that helps. Yeah, that makes sense because also to make a proper workflow, you do have to be a big company to ship it and make sure it’s like you and to eval it.
50:24 Because right now all the skills, it’s really expensive to eval the skill on all the double traces where the skill would be used.
50:31 And that’s for individual skill. And to eval it workflow with a bunch of these skills is like an order of magnitude higher amount of effort.
50:41 So yeah, no doubt there’s… Oh yeah, I mean, tell me, like, I feel like I’m partially going insane from it.
50:49 But I do actually, like, my personal thesis on what I think would be very, very kind of a hot space is the idea of how to maximize legibility of what the agent, like, the
51:03 swarm doing while still maintaining the same speed of just giving one prompt.
51:16 It always comes back to evals for how well these tools work.
51:21 It’s the hardest, most tedious thing to do, but also the most useful for all of it.
51:30 Yeah. Anybody else have other thoughts about the harnesses, how they work, skills, anything like that in the last three minutes?
51:42 I think there are some cool ones like MIT in another tone.
51:47 MIT Nanaton is very good. I create several projects with it.
51:55 Okay, just in terms of just like a harness and flow to use?
51:59 Collaboration, all of the agents are collaborate with each other.
52:03 It’s very easy to use. I know we actually, a decent number of people here have worked on Nanda at some point.
52:13 So is this sort of like a codex replacement kind of thing, or like a clawed code replacement?
52:20 Multi-agents work. Easily work. Okay.
52:33 Or the last two minutes. Any other things you want to call out?
52:38 I have a piece of news that I like, you know, very unremarkable, that GPT Terra is getting retired.
52:48 And I just discovered GPT-Terra is like a as a good golden spot for, because I found Luna to be very, very like fine-tuned on like to not probably not
53:02 fine-tune, probably more like RL, but to reject and be like, I need authorization.
53:06 I had like literally some like workflow once where it needed to go to Amazon, confirm it’s in the right account, and like put something in the card.
53:14 And instead of like, and the Amazon, instead of saying, hi, Daniel, like it should have been in the, and it’s mentioned in the skill, it said, hello, Daniel.
53:23 And it was like, no, it’s not the right account.
53:25 I can’t do it. I’m not authorized. And Luna kept doing things, like stupid things like that.
53:30 And I found Terra is sufficiently cheap to not burn as much credits as Saul for like silly things and also is not annoyingly stubborn to not require authorization.
53:42 And they’re killing it already. So it’s just, I’ve been using it for like a week at this point, had more credits to spare for things like this that don’t require like full knowledge, but need to not be that picky.
53:58 But it’s still hidden around for a while, right?
54:01 It’s yeah. It’s yeah, definitely, yeah.
54:03 As models get smaller, they get more brittle in terms of their instruction following, where things need to be more precise.
54:09 It’s, I believe, the reason they’re Sunsetting Terra is because they’re releasing the GPT-6 versions of them.
54:17 I’ve heard reports of people getting routed to GPT-6 Seoul in the past day or so.
54:24 So people are expecting the smaller models for the GPT-6 series to go and come out.
54:31 And that Astra will probably be the biggest, but we’ll still have a sole Terra and Luna, just GPT-6 class instead of GPT-5.6.
54:39 So I think that might be why. So you might be getting your Terra back soon, or an upgraded version of it.
54:46 Or they might re because they would have like four tiers because Astra is a tier higher than Sol, right?
54:50 So there would be like four tiers of models.
54:53 So I guess they did similar to Anthropic, suddenly.
54:56 And I guess they might just rebrand Luna as Terra or something to not have a…
55:01 like because it’s like named after planets would be a bit of a big jump from going from the moon to the sun but uh yeah maybe they’ll prepare because i probably it does feel like a strange like arrangement
55:16 to have four tiers right like you’d have a middle low etc why would you use it otherwise i also feel like like like astra’s not that much smarter than soul right like wouldn’t a gbt gbt6
55:30 soul actually be smarter than astra it’s i think astra is definitely the biggest model i think for a lot of tasks it’s just getting very hard to measure the difference between these models uh because i think most of the sort
55:43 of like public benchmarks are fairly saturated um i mean like gpt6 is very obviously um much better than gpt 5.6 at image understanding and also computer use um but
55:57 i think coding for most coding stuff we’re already sort of like near the limit of like how good we can you know write code they both write code correctly you know like they both get the feature done and a lot of it now is
56:08 coming down to like style and efficiency uh for a lot of the mundane tasks um this is why i say sort of like in the future we’re not going to need these frontier models for most things because they’re going to be indistinguishable
56:18 from like the mid-tier and even smaller tier models so yeah i think astero will still have a place i think soul will be a bit worse uh in probably some dimensions that we can you know easily see in the real world um but yeah
56:35 in the future just anthropic through deep sea yeah uh cool and with that we have reached the end of our ai tools club session today uh
56:49 thank you all for coming and i hope to see you all next week
Subscribe to get the latest AI news in your inbox every week!
