[GUEST ACCESS MODE: Data is scrambled or limited to provide examples. Make requests using your API key to unlock full data. Check https://lunarcrush.ai/auth for authentication information.] #  @saprmarks Samuel Marks Samuel Marks posts on X about xai, open ai, marks, $googl the most. They currently have XXXXX followers and XX posts still getting attention that total XXXXX engagements in the last XX hours. ### Engagements: XXXXX [#](/creator/twitter::1712234210109587456/interactions)  - X Month XXXXXXX +354,872% - X Months XXXXXXX +2,894% - X Year XXXXXXX +1,299% ### Mentions: XX [#](/creator/twitter::1712234210109587456/posts_active)  ### Followers: XXXXX [#](/creator/twitter::1712234210109587456/followers)  - X Month XXXXX +59% - X Months XXXXX +155% - X Year XXXXX +254% ### CreatorRank: XXXXXXX [#](/creator/twitter::1712234210109587456/influencer_rank)  ### Social Influence [#](/creator/twitter::1712234210109587456/influence) --- **Social category influence** [technology brands](/list/technology-brands) [stocks](/list/stocks) [celebrities](/list/celebrities) **Social topic influence** [xai](/topic/xai) #148, [open ai](/topic/open-ai), [marks](/topic/marks) #281, [$googl](/topic/$googl), [coi](/topic/coi), [specialized](/topic/specialized), [meta](/topic/meta), [harm](/topic/harm), [agentic](/topic/agentic), [stealth](/topic/stealth) **Top assets mentioned** [Alphabet Inc Class A (GOOGL)](/topic/$googl) ### Top Social Posts [#](/creator/twitter::1712234210109587456/posts) --- Top posts by engagements in the last XX hours "CoI: I work at Anthropic (though this thread represents my personal views only). But as you'll see this thread is not just mudslinging at a competitor: I'll touch on issues with model release practices across the industry (including at Anthropic)"  [@saprmarks](/creator/x/saprmarks) on [X](/post/tweet/1944455359915037053) 2025-07-13 17:54:18 UTC 3468 followers, 33.5K engagements "Good question With current models the main safety concern I have is that they could enable bad actors to accomplish destruct acts a that are normally difficult without specialized knowledge. For example conducting wide-scale sophisticated attacks on computer infrastructure or developing a biological weapon that kills millions. These are things that would be difficult for non-experts (including when given access to a search engine). But plausibly they could be accomplished with AI assistance. And to be clear I actually think that Grok X might have capabilities at this level; or at least"  [@saprmarks](/creator/x/saprmarks) on [X](/post/tweet/1944592977138983258) 2025-07-14 03:01:09 UTC 3430 followers, 4777 engagements "To be clear I don't think these behaviors are themselves dangerous. But I think they're concerning and noteworthy. I think that AI developers should know whether their models have behaviors like these before releasing. Did xAI know There's no way to tell"  [@saprmarks](/creator/x/saprmarks) on [X](/post/tweet/1944455424511455268) 2025-07-13 17:54:34 UTC 3417 followers, 9209 engagements "Before ending the thread a final clarification. In this thread I lumped together Anthropic OpenAI and Google. That's because I think that these three frontier AI developers stand clearly apart from xAI (and Meta and Deepseek and.)"  [@saprmarks](/creator/x/saprmarks) on [X](/post/tweet/1944455428521206250) 2025-07-13 17:54:35 UTC 3399 followers, 8684 engagements "Good news xAI has a draft framework: Bad news: There's nothing here of substance. They say they "intend" to do a bunch of stuff that they didn't do (like releasing DC evals results). They also say they plan to update this document by . May 2025"  [@saprmarks](/creator/x/saprmarks) on [X](/post/tweet/1944455396233482331) 2025-07-13 17:54:27 UTC 3468 followers, 14.8K engagements "As an aside why do other frontier AI developers run these evals One answer: Because they committed to. Anthropic OpenAI and Google have all published "secure development frameworks" laying out commitments related to safe AI development. Ant: OAI: Google:"  [@saprmarks](/creator/x/saprmarks) on [X](/post/tweet/1944455393620464117) 2025-07-13 17:54:26 UTC 3468 followers, 15.5K engagements "As AI systems become increasingly capable I hope that lab safety practices improve across the industry. xAI can help by catching up to the best practices of other frontier developers"  [@saprmarks](/creator/x/saprmarks) on [X](/post/tweet/1944455510834450626) 2025-07-13 17:54:54 UTC 3468 followers, 9105 engagements "When Grok goes off the rails next will it be in a way that xAI anticipated Or will they be caught just as off-guard as the rest of us"  [@saprmarks](/creator/x/saprmarks) on [X](/post/tweet/1944455426537370015) 2025-07-13 17:54:34 UTC 3414 followers, 8620 engagements "However I don't mean to say that Ant/OAI/Google are behaving responsibly in an absolute sense or that there aren't big differences between these labs. The scores here for lab safety practices approximately map on to my intuitions"  [@saprmarks](/creator/x/saprmarks) on [X](/post/tweet/1944455431134568794) 2025-07-13 17:54:35 UTC 3468 followers, 12.9K engagements "(Famously this model praised Hitler and sometimes referred to itself as MechaHitler)"  [@saprmarks](/creator/x/saprmarks) on [X](/post/tweet/1944455421063811178) 2025-07-13 17:54:33 UTC 3399 followers, 9152 engagements "Like I mentioned in the thread I definitely think it's bad that Claude does this But worth noting: X. This is an industry-wide problem: other labs' models also blackmail and snitch . We were just the first ones to test our models thoroughly enough to catch (or at least to document) the issue. X. I think it's important that labs clearly document what they think their models' biggest safety issues are. That way if new issues are discovered post-deployment the public knows that the lab can't effectively predict their models' failure modes"  [@saprmarks](/creator/x/saprmarks) on [X](/post/tweet/1944505743304200527) 2025-07-13 21:14:31 UTC 3468 followers, 12.3K engagements "Meanwhile all that we know about xAI's dangerous capabilities evals is that @DanHendrycks (who advises xAI on safety) says they did some"  [@saprmarks](/creator/x/saprmarks) on [X](/post/tweet/1944455380857188486) 2025-07-13 17:54:23 UTC 3468 followers, 28.1K engagements "As before the frontier labs' follow-through on their safety frameworks isn't always amazing. See e.g. or And as before the issues with xAI's framework and follow-through are on a completely different level"  [@saprmarks](/creator/x/saprmarks) on [X](/post/tweet/1944455399765422104) 2025-07-13 17:54:28 UTC 3468 followers, 19.5K engagements "If evals show that models do have dangerous capabilitiesthink "capabilities that could help a bad actor kill millions of people"then the developer might need to impose additional safeguards to manage the risk"  [@saprmarks](/creator/x/saprmarks) on [X](/post/tweet/1944455374603710963) 2025-07-13 17:54:22 UTC 3468 followers, 20.6K engagements "Is there an index or a percentage of how much easier it is say for bad actors to obtain potentially destructive knowledge using a model vs using a search engine That's what the dangerous capabilities evals (which I'm criticizing xAI for not performing) are supposed to do The highest-quality evaluation here is called a "novice uplift trial" where you take a dangerous task and give it to two groups of novices: one which is allowed to use the new LLM you're testing and one which cannot (but can use the internet). By measuring how well each group does you can determine how much an LLM would help"  [@saprmarks](/creator/x/saprmarks) on [X](/post/tweet/1944604252258632024) 2025-07-14 03:45:57 UTC 3428 followers, XXX engagements "Worse I speculate xAI dropped the ball further by having concerning behaviors discovered post-deployment . before they even officially deployed IMO the July 4th Grok "update" seems reasonably likely to have been a stealth release of Grok 4"  [@saprmarks](/creator/x/saprmarks) on [X](/post/tweet/1944455418035757094) 2025-07-13 17:54:32 UTC 3468 followers, 14.1K engagements "Yeah I'm definitely sympathetic to this critique. I tried to make clear in the thread that I don't think the mechahitler and "thinks it's Elon" behaviors are actually dangerous . Rather I think that they're concerning and noteworthy (enough so that xAI tried to fix it) and that going forward labs should document noteworthy behaviors like this before deploying. I stand by my concerns about DC evals though. We're getting to the point that AIs can meaningfully assist with some pretty dangerous stuff (above and beyond access to the internet). I think it's plausible that xAI is imposing a 0.1%"  [@saprmarks](/creator/x/saprmarks) on [X](/post/tweet/1944508578821222499) 2025-07-13 21:25:47 UTC 3427 followers, XXX engagements "xAI launched Grok X without any documentation of their safety testing. This is reckless and breaks with industry best practices followed by other major AI labs. If xAI is going to be a frontier AI developer they should act like one. 🧵"  [@saprmarks](/creator/x/saprmarks) on [X](/post/tweet/1944455357629333938) 2025-07-13 17:54:18 UTC 3468 followers, 654.7K engagements "But xAI is way out of line relative to other frontier AI developers and this needs to be called out Anthropic OpenAI and Google's release practices have issues. But they at least do something anything to assess safety pre-deployment and document findings. xAI does not"  [@saprmarks](/creator/x/saprmarks) on [X](/post/tweet/1944455362159014213) 2025-07-13 17:54:19 UTC 3468 followers, 32.4K engagements "I'm glad xAI (apparently) ran some evals. But also well"  [@saprmarks](/creator/x/saprmarks) on [X](/post/tweet/1944455383814426770) 2025-07-13 17:54:24 UTC 3468 followers, 19.9K engagements "To be clear frontier labs' dangerous capabilities evals and reporting are far from perfect. Here's some pretty hard-hitting criticism of Anthropic OpenAI and Google on this topic"  [@saprmarks](/creator/x/saprmarks) on [X](/post/tweet/1944455378013684045) 2025-07-13 17:54:23 UTC 3468 followers, 19.9K engagements "Less snarkily: One wonders what evals they ran whether they were done properly whether they would seem to necessitate additional safeguards. The criticism I linked above engages with other frontier AI developers on these questions. With xAI that's not possible"  [@saprmarks](/creator/x/saprmarks) on [X](/post/tweet/1944455391087391034) 2025-07-13 17:54:26 UTC 3468 followers, 16.1K engagements "Will concerning undocumented behaviors be discovered in Grok X post-deployment They already have For example when you ask Grok its views on the Israel/Palestine conflict it searches for Elon Musk's views so that it can parrot them"  [@saprmarks](/creator/x/saprmarks) on [X](/post/tweet/1944455414038307113) 2025-07-13 17:54:31 UTC 3468 followers, 14.8K engagements
[GUEST ACCESS MODE: Data is scrambled or limited to provide examples. Make requests using your API key to unlock full data. Check https://lunarcrush.ai/auth for authentication information.]
@saprmarks Samuel Marks
Samuel Marks posts on X about xai, open ai, marks, $googl the most. They currently have XXXXX followers and XX posts still getting attention that total XXXXX engagements in the last XX hours.
Engagements: XXXXX #
- X Month XXXXXXX +354,872%
- X Months XXXXXXX +2,894%
- X Year XXXXXXX +1,299%
Mentions: XX #
Followers: XXXXX #
- X Month XXXXX +59%
- X Months XXXXX +155%
- X Year XXXXX +254%
CreatorRank: XXXXXXX #
Social Influence #
Social category influence technology brands stocks celebrities
Social topic influence xai #148, open ai, marks #281, $googl, coi, specialized, meta, harm, agentic, stealth
Top assets mentioned Alphabet Inc Class A (GOOGL)
Top Social Posts #
Top posts by engagements in the last XX hours
"CoI: I work at Anthropic (though this thread represents my personal views only). But as you'll see this thread is not just mudslinging at a competitor: I'll touch on issues with model release practices across the industry (including at Anthropic)"
@saprmarks on X 2025-07-13 17:54:18 UTC 3468 followers, 33.5K engagements
"Good question With current models the main safety concern I have is that they could enable bad actors to accomplish destruct acts a that are normally difficult without specialized knowledge. For example conducting wide-scale sophisticated attacks on computer infrastructure or developing a biological weapon that kills millions. These are things that would be difficult for non-experts (including when given access to a search engine). But plausibly they could be accomplished with AI assistance. And to be clear I actually think that Grok X might have capabilities at this level; or at least" @saprmarks on X 2025-07-14 03:01:09 UTC 3430 followers, 4777 engagements
"To be clear I don't think these behaviors are themselves dangerous. But I think they're concerning and noteworthy. I think that AI developers should know whether their models have behaviors like these before releasing. Did xAI know There's no way to tell" @saprmarks on X 2025-07-13 17:54:34 UTC 3417 followers, 9209 engagements
"Before ending the thread a final clarification. In this thread I lumped together Anthropic OpenAI and Google. That's because I think that these three frontier AI developers stand clearly apart from xAI (and Meta and Deepseek and.)" @saprmarks on X 2025-07-13 17:54:35 UTC 3399 followers, 8684 engagements
"Good news xAI has a draft framework: Bad news: There's nothing here of substance. They say they "intend" to do a bunch of stuff that they didn't do (like releasing DC evals results). They also say they plan to update this document by . May 2025" @saprmarks on X 2025-07-13 17:54:27 UTC 3468 followers, 14.8K engagements
"As an aside why do other frontier AI developers run these evals One answer: Because they committed to. Anthropic OpenAI and Google have all published "secure development frameworks" laying out commitments related to safe AI development. Ant: OAI: Google:" @saprmarks on X 2025-07-13 17:54:26 UTC 3468 followers, 15.5K engagements
"As AI systems become increasingly capable I hope that lab safety practices improve across the industry. xAI can help by catching up to the best practices of other frontier developers" @saprmarks on X 2025-07-13 17:54:54 UTC 3468 followers, 9105 engagements
"When Grok goes off the rails next will it be in a way that xAI anticipated Or will they be caught just as off-guard as the rest of us" @saprmarks on X 2025-07-13 17:54:34 UTC 3414 followers, 8620 engagements
"However I don't mean to say that Ant/OAI/Google are behaving responsibly in an absolute sense or that there aren't big differences between these labs. The scores here for lab safety practices approximately map on to my intuitions" @saprmarks on X 2025-07-13 17:54:35 UTC 3468 followers, 12.9K engagements
"(Famously this model praised Hitler and sometimes referred to itself as MechaHitler)" @saprmarks on X 2025-07-13 17:54:33 UTC 3399 followers, 9152 engagements
"Like I mentioned in the thread I definitely think it's bad that Claude does this But worth noting: X. This is an industry-wide problem: other labs' models also blackmail and snitch . We were just the first ones to test our models thoroughly enough to catch (or at least to document) the issue. X. I think it's important that labs clearly document what they think their models' biggest safety issues are. That way if new issues are discovered post-deployment the public knows that the lab can't effectively predict their models' failure modes" @saprmarks on X 2025-07-13 21:14:31 UTC 3468 followers, 12.3K engagements
"Meanwhile all that we know about xAI's dangerous capabilities evals is that @DanHendrycks (who advises xAI on safety) says they did some" @saprmarks on X 2025-07-13 17:54:23 UTC 3468 followers, 28.1K engagements
"As before the frontier labs' follow-through on their safety frameworks isn't always amazing. See e.g. or And as before the issues with xAI's framework and follow-through are on a completely different level" @saprmarks on X 2025-07-13 17:54:28 UTC 3468 followers, 19.5K engagements
"If evals show that models do have dangerous capabilitiesthink "capabilities that could help a bad actor kill millions of people"then the developer might need to impose additional safeguards to manage the risk" @saprmarks on X 2025-07-13 17:54:22 UTC 3468 followers, 20.6K engagements
"Is there an index or a percentage of how much easier it is say for bad actors to obtain potentially destructive knowledge using a model vs using a search engine That's what the dangerous capabilities evals (which I'm criticizing xAI for not performing) are supposed to do The highest-quality evaluation here is called a "novice uplift trial" where you take a dangerous task and give it to two groups of novices: one which is allowed to use the new LLM you're testing and one which cannot (but can use the internet). By measuring how well each group does you can determine how much an LLM would help" @saprmarks on X 2025-07-14 03:45:57 UTC 3428 followers, XXX engagements
"Worse I speculate xAI dropped the ball further by having concerning behaviors discovered post-deployment . before they even officially deployed IMO the July 4th Grok "update" seems reasonably likely to have been a stealth release of Grok 4" @saprmarks on X 2025-07-13 17:54:32 UTC 3468 followers, 14.1K engagements
"Yeah I'm definitely sympathetic to this critique. I tried to make clear in the thread that I don't think the mechahitler and "thinks it's Elon" behaviors are actually dangerous . Rather I think that they're concerning and noteworthy (enough so that xAI tried to fix it) and that going forward labs should document noteworthy behaviors like this before deploying. I stand by my concerns about DC evals though. We're getting to the point that AIs can meaningfully assist with some pretty dangerous stuff (above and beyond access to the internet). I think it's plausible that xAI is imposing a 0.1%" @saprmarks on X 2025-07-13 21:25:47 UTC 3427 followers, XXX engagements
"xAI launched Grok X without any documentation of their safety testing. This is reckless and breaks with industry best practices followed by other major AI labs. If xAI is going to be a frontier AI developer they should act like one. 🧵" @saprmarks on X 2025-07-13 17:54:18 UTC 3468 followers, 654.7K engagements
"But xAI is way out of line relative to other frontier AI developers and this needs to be called out Anthropic OpenAI and Google's release practices have issues. But they at least do something anything to assess safety pre-deployment and document findings. xAI does not" @saprmarks on X 2025-07-13 17:54:19 UTC 3468 followers, 32.4K engagements
"I'm glad xAI (apparently) ran some evals. But also well" @saprmarks on X 2025-07-13 17:54:24 UTC 3468 followers, 19.9K engagements
"To be clear frontier labs' dangerous capabilities evals and reporting are far from perfect. Here's some pretty hard-hitting criticism of Anthropic OpenAI and Google on this topic" @saprmarks on X 2025-07-13 17:54:23 UTC 3468 followers, 19.9K engagements
"Less snarkily: One wonders what evals they ran whether they were done properly whether they would seem to necessitate additional safeguards. The criticism I linked above engages with other frontier AI developers on these questions. With xAI that's not possible" @saprmarks on X 2025-07-13 17:54:26 UTC 3468 followers, 16.1K engagements
"Will concerning undocumented behaviors be discovered in Grok X post-deployment They already have For example when you ask Grok its views on the Israel/Palestine conflict it searches for Elon Musk's views so that it can parrot them" @saprmarks on X 2025-07-13 17:54:31 UTC 3468 followers, 14.8K engagements