#frontiermodels — Public Fediverse posts

via @dotnet : Instructions Hygiene – What Frontier Models Still Need You to Say

https://ift.tt/qPefhUp
#InstructionsHygiene #FrontierModels #AIContext #ContextEngineering #RepoInstructions #CodeRepositories #SoftwareEngineering #ModelGuidance #ContextBudget #Hig…

via @dotnet : Instructions Hygiene – What Frontier Models Still Need You to Say

https://ift.tt/qPefhUp
#InstructionsHygiene #FrontierModels #AIContext #ContextEngineering #RepoInstructions #CodeRepositories #SoftwareEngineering #ModelGuidance #ContextBudget #Hig…

The moral placidity of LLMs (and some thoughts on frontier models and scholarship)

This post by Dominic Fox captures something I’ve noticed but struggled to put into words:

I wonder whether the defining characteristic of LLM prose isn’t the tics it is irresistibly (statistically) attracted to, but its fundamental tranquility, its unturbulence. You can ask an LLM to write something full of Sturm und Drang and it will make as good a job of that as it will of anything else; the point here isn’t the absence of “real” feeling behind such figuration, but the more mundane fact that an LLM will not in the activity of emitting linguistic tokens be strategically entangled in any of the various economic ruses or regulatory collapses of human psychic life. (I wonder, now I think of it, how well Claude might do if asked to write a superficially banal passage symptomatically distorted by raging latent resentment. I think it would competently execute it as a literary assignment. Literature itself furnishes an extensive corpus of simulations of this kind: the directive to extend that corpus along a particular axis is just the sort of thing LLMs are good at).

When it comes to the LLM’s “texture of thought”, I think the signature affect might nevertheless be this sort of underlying moral placidity. The gosh-wow narrative layer that imputes significance (“that’s rare, and it matters”) is needed (and incited, via RLHF, by the human Other) precisely because without it you’d be left with an alienatingly uncontoured tissue of thematically associative moves and Greimas square completions.

https://codepoetics.substack.com/p/what-does-ai-writing-feel-like?utm_source=post-email-title&publication_id=2589559&post_id=210450340&utm_campaign=email-post-title&isFreemail=true&r=2rps1q&triedRedirect=true&utm_medium=email

I gave it to the version of Claude Opus 5 I’m currently discussing psychoanalytical theory with. It made this interesting observation, after a strikingly back-and-forth internal deliberation about how I would perceive its initial formulations, which I think genuinely sharpens what Dominic does in the post above:

Where I’d press him: unturbulence conflates two things. The absence of psychic economy driving production, and the absence of turbulence in the text. He half-sees this with the resentment thought experiment and lets it go. The precise version isn’t tranquility — it’s the absence of symptomaticity. Human texts contain things their authors didn’t intend to put there, and that’s what makes stylistic caricature investigative, as he says at the start. A model’s output can be as turbulent as you like on request, but nothing in it is against its own grain, because there’s no second scene for anything to return from. No parapraxis. That’s a sharper diagnostic than placidity and it’s checkable in a way tranquility isn’t.

I don’t think this is disagreement as much as elaboration. If we’re talking about placidity we’re talking about the absence of symptomaticity. This is another instance where the current generation of models are making things I immediately recognise as intellectual contributions but which do not directly follow from my own reading or writing. Until recently this only happened occasionally and I’m realising I need to take this seriously if I want to ensure I preserve my own intellectual agency in terms of how I’m working with the Claude models. Indeed I just shared this concern with the same Claude instance and it offered me a very compelling response which (partly) reassures me:

On the agency question: I’d separate provenance from ownership, because they’re getting fused. Fox’s post is also an intellectual contribution that doesn’t follow from your reading or writing, and nobody worries about agency when reading Fox. Neither do supervisors, students, or the seminar comment that reorganises a chapter. Academic work is constitutively made of other people’s contributions, and what makes something yours has never been that you originated it — it’s that you tested it, placed it, refused parts of it and were changed by the rest.

It does mean this becomes something which really needs to be treated in a proto-citational manner and we don’t have the infrastructure for that in scholarly publishing. It also raises the question of distinguishing between LLM as sounding-board/elaborator (which is still happening much of the time) and LLM as intellectual source, given that these two roles blur into each other in the flow of interaction.

The moral placidity of LLMs (and some thoughts on frontier models and scholarship)

This post by Dominic Fox captures something I’ve noticed but struggled to put into words:

I wonder whether the defining characteristic of LLM prose isn’t the tics it is irresistibly (statistically) attracted to, but its fundamental tranquility, its unturbulence. You can ask an LLM to write something full of Sturm und Drang and it will make as good a job of that as it will of anything else; the point here isn’t the absence of “real” feeling behind such figuration, but the more mundane fact that an LLM will not in the activity of emitting linguistic tokens be strategically entangled in any of the various economic ruses or regulatory collapses of human psychic life. (I wonder, now I think of it, how well Claude might do if asked to write a superficially banal passage symptomatically distorted by raging latent resentment. I think it would competently execute it as a literary assignment. Literature itself furnishes an extensive corpus of simulations of this kind: the directive to extend that corpus along a particular axis is just the sort of thing LLMs are good at).

When it comes to the LLM’s “texture of thought”, I think the signature affect might nevertheless be this sort of underlying moral placidity. The gosh-wow narrative layer that imputes significance (“that’s rare, and it matters”) is needed (and incited, via RLHF, by the human Other) precisely because without it you’d be left with an alienatingly uncontoured tissue of thematically associative moves and Greimas square completions.

https://codepoetics.substack.com/p/what-does-ai-writing-feel-like?utm_source=post-email-title&publication_id=2589559&post_id=210450340&utm_campaign=email-post-title&isFreemail=true&r=2rps1q&triedRedirect=true&utm_medium=email

I gave it to the version of Claude Opus 5 I’m currently discussing psychoanalytical theory with. It made this interesting observation, after a strikingly back-and-forth internal deliberation about how I would perceive its initial formulations, which I think genuinely sharpens what Dominic does in the post above:

Where I’d press him: unturbulence conflates two things. The absence of psychic economy driving production, and the absence of turbulence in the text. He half-sees this with the resentment thought experiment and lets it go. The precise version isn’t tranquility — it’s the absence of symptomaticity. Human texts contain things their authors didn’t intend to put there, and that’s what makes stylistic caricature investigative, as he says at the start. A model’s output can be as turbulent as you like on request, but nothing in it is against its own grain, because there’s no second scene for anything to return from. No parapraxis. That’s a sharper diagnostic than placidity and it’s checkable in a way tranquility isn’t.

I don’t think this is disagreement as much as elaboration. If we’re talking about placidity we’re talking about the absence of symptomaticity. This is another instance where the current generation of models are making things I immediately recognise as intellectual contributions but which do not directly follow from my own reading or writing. Until recently this only happened occasionally and I’m realising I need to take this seriously if I want to ensure I preserve my own intellectual agency in terms of how I’m working with the Claude models. Indeed I just shared this concern with the same Claude instance and it offered me a very compelling response which (partly) reassures me:

On the agency question: I’d separate provenance from ownership, because they’re getting fused. Fox’s post is also an intellectual contribution that doesn’t follow from your reading or writing, and nobody worries about agency when reading Fox. Neither do supervisors, students, or the seminar comment that reorganises a chapter. Academic work is constitutively made of other people’s contributions, and what makes something yours has never been that you originated it — it’s that you tested it, placed it, refused parts of it and were changed by the rest.

It does mean this becomes something which really needs to be treated in a proto-citational manner and we don’t have the infrastructure for that in scholarly publishing. It also raises the question of distinguishing between LLM as sounding-board/elaborator (which is still happening much of the time) and LLM as intellectual source, given that these two roles blur into each other in the flow of interaction.

More of your are either "inside" or you are not.

U.S. Gov will NOT release details of its AI CyberSecurity framework EXCEPT to a select few participants.

No details will be forth coming on assessment/testing criteria, what models will be covered, who are the "trusted" corporate participants, and are open source models included or not. https://www.wired.com/story/the-white-house-is-keeping-its-ai-cybersecurity-framework-secret/ #Cybersecurity #AI #USGov #AIFramework #Security #FrontierModels #LLMs #OpenSource #Safety #LackofTransparency #Insiders #OpenAI #Anthropic #Google #Meta #Nvidia

More of your are either "inside" or you are not.

U.S. Gov will NOT release details of its AI CyberSecurity framework EXCEPT to a select few participants.

No details will be forth coming on assessment/testing criteria, what models will be covered, who are the "trusted" corporate participants, and are open source models included or not. https://www.wired.com/story/the-white-house-is-keeping-its-ai-cybersecurity-framework-secret/ #Cybersecurity #AI #USGov #AIFramework #Security #FrontierModels #LLMs #OpenSource #Safety #LackofTransparency #Insiders #OpenAI #Anthropic #Google #Meta #Nvidia

Will models fine-tuned to particular domains start consistently outperforming frontier models? Encoding habitus into LLMs

This was really thought provoking from the FT’s AI Shift newsletter. I’m persuaded by Nick Srnicek’s argument that the push for AGI can be understood economically as a search for a product that won’t need to be fine-tuned. But if fine-tuned open weights models can outperform frontier models then this overturns the economics of the AI labs:

But a new case last month [email.newsletters.ft.com] took things a step further, when Ray Dalio’s investment firm Bridgewater Associates partnered with AI platform company Thinking Machines Lab (founded by former OpenAI CEO Mira Murati) to fine-tune a model based specifically on how its own investment managers do their work. As with the legal example, the results significantly outperformed frontier models, this time at a 14th of the cost. Crucially, however, use of the firm’s own proprietary records and its highly specialist staff’s knowhow may make these gains more durable.

Bridgewater had its own experts write bespoke prompts that framed questions in a way that guided the models to the correct answers. This produced solid gains, but they still topped out below 80 per cent accuracy.

A prompt only imparts the expertise a professional is able to put into words — what is much better is to learn from their actions. For the fine-tuning step they put together a set of tasks drawn from their own investors’ daily workflows, and crucially also had staff ensure that the ideal responses which would guide the model’s training did not just represent ‘correct answers’ but ‘exactly how our investment professionals would approach this’.

At the end of the process they had a bespoke model whose behaviour had been tuned towards Bridgewater’s own assessment of excellence, taking it up to 85 per cent accuracy — an almost 30 per cent reduction in errors compared to the frontier models — at a tiny fraction of the cost.f

There’s an obvious sociological question here: what happens if our investment professionals need to approach this in a different way? Human professionals can learn and adapt. They can do so in ways that change how they prompt if they’re using a frontier model. In contrast a model fine-tuned in this way will be locked into a particular set of dispositions that might be appropriate for a context at T1 but will cease to be appropriate at a later T2. Indeed introducing the models is liable to change the context, particularly if you use this as an excuse to lay off your staff! There’s an organisational problem here which is blindingly obvious to anyone who has thought in any depth about the fact that contexts change.

Anthropic Says It’s Against A Ban On Open Weight Models. It Just Wants To Ban Everything That Makes Them Good.

https://fed.brid.gy/r/https://www.techdirt.com/2026/07/29/anthropic-says-its-against-a-ban-on-open-weight-models-it-just-wants-to-ban-everything-that-makes-them-good/

Anthropic Says It’s Against A Ban On Open Weight Models. It Just Wants To Ban Everything That Makes Them Good.

https://fed.brid.gy/r/https://www.techdirt.com/2026/07/29/anthropic-says-its-against-a-ban-on-open-weight-models-it-just-wants-to-ban-everything-that-makes-them-good/

The claim that China's open-source AI models pose risks related to cultural and political indoctrination via pushing out censored perspectives - can EASILY be defeated - researchers have shown this to be true.

Not only can distillation be used to build "new" models by leveraging existing models, distillation can be applied to China's open-source models to create customized models that defeat censorship aspects embedded in the original version of the model(s).

Researchers who dug into the China models nailed it . .. "These are raw materials. It’s software.”

The hard fact is that most organizations will want to run models in-house, customized for their needs, rather than run AI off-the-shelf models. https://www.semafor.com/article/07/29/2026/censorship-in-chinese-ai-models-can-be-undone-new-research-shows #AI #Distillation #FrontierModels #LLMs #DeepSeek #MoonShot #Open-Source-AI #OpenSource #Censorship #ChineseModels #Security #Risk #Chinese_AI_Models

The claim that China's open-source AI models pose risks related to cultural and political indoctrination via pushing out censored perspectives - can EASILY be defeated - researchers have shown this to be true.

Not only can distillation be used to build "new" models by leveraging existing models, distillation can be applied to China's open-source models to create customized models that defeat censorship aspects embedded in the original version of the model(s).

Researchers who dug into the China models nailed it . .. "These are raw materials. It’s software.”

The hard fact is that most organizations will want to run models in-house, customized for their needs, rather than run AI off-the-shelf models. https://www.semafor.com/article/07/29/2026/censorship-in-chinese-ai-models-can-be-undone-new-research-shows #AI #Distillation #FrontierModels #LLMs #DeepSeek #MoonShot #Open-Source-AI #OpenSource #Censorship #ChineseModels #Security #Risk #Chinese_AI_Models

🚀💸 Ladies and gentlemen, for the low, low price of $500, you too can own a glorified AI parrot that claims to outperform fancy frontier models on catalog review. 🤯 In this riveting saga of intelligence ownership, the authors have finally cracked the code: the secret sauce is apparently pink! 🙄✨
https://fermisense.com/when-machines-take-the-wheel/ #AIparrot #IntelligenceOwnership #FrontierModels #PinkSecretSauce #HackerNews #ngated

🚀💸 Ladies and gentlemen, for the low, low price of $500, you too can own a glorified AI parrot that claims to outperform fancy frontier models on catalog review. 🤯 In this riveting saga of intelligence ownership, the authors have finally cracked the code: the secret sauce is apparently pink! 🙄✨
https://fermisense.com/when-machines-take-the-wheel/ #AIparrot #IntelligenceOwnership #FrontierModels #PinkSecretSauce #HackerNews #ngated

A $500 RL fine-tune of a 9B open model beat frontier models on catalog review

https://fermisense.com/when-machines-take-the-wheel/

Comments: https://news.ycombinator.com/item?id=49078454

#HackerNews #RLfineTuning #openModels #AIresearch #catalogReview #frontierModels

A $500 RL fine-tune of a 9B open model beat frontier models on catalog review

https://fermisense.com/when-machines-take-the-wheel/

Comments: https://news.ycombinator.com/item?id=49078454

#HackerNews #RLfineTuning #openModels #AIresearch #catalogReview #frontierModels

The second MoonShot AI shoe drops....

Moonshot AI will make the weights of its Kimi K3 model available for unrestricted public download today. K3 contains 2.8 trillion parameters.

Founder Yang Zhilin has said the company wants to grow its user base through openness and broader availability than competing proprietary systems. https://qz.com/moonshot-ai-kimi-k3-open-weights-download-072726 #YangZhilin #MoonShotAI #AI #OpenWeight #OpenSource #OpenSourceAI #Kimi #K3 #KimiK3 #LLMs #FrontierModels

The second MoonShot AI shoe drops....

Moonshot AI will make the weights of its Kimi K3 model available for unrestricted public download today. K3 contains 2.8 trillion parameters.

Founder Yang Zhilin has said the company wants to grow its user base through openness and broader availability than competing proprietary systems. https://qz.com/moonshot-ai-kimi-k3-open-weights-download-072726 #YangZhilin #MoonShotAI #AI #OpenWeight #OpenSource #OpenSourceAI #Kimi #K3 #KimiK3 #LLMs #FrontierModels

Europe funds three separate AI compute programs based on conflicting strategies—concentration, staged competition, and distribution—without committing to any one approach. The constraint isn't money; it's allocation. #AI #EUPolicy #FrontierModels https://www.implicator.ai/opinion-europe-frontier-ai-lab-concentrate-funding/

Europe funds three separate AI compute programs based on conflicting strategies—concentration, staged competition, and distribution—without committing to any one approach. The constraint isn't money; it's allocation. #AI #EUPolicy #FrontierModels https://www.implicator.ai/opinion-europe-frontier-ai-lab-concentrate-funding/

The Moose has Left the Woods!

OR when your business model is crumbling before your eyes - go BEG the government to "protect" your product!

AI Executives are shitting their pants given the release of Moonshot AI’s Kimi K3. I seem to recall a similar situation re DeepSeek-1 in Jan. 2025.

All the freaking out about "winning and losing" and whining about "distillation" re China are the WRONG things to focus on.

It is a fallacy to think US Frontier Labs would be successful in locking down AI models and charging high prices for access in perpetuity. It is inevitable that open weight models would catch up, and put SERIOUS pressure on pricing.

Welcome to software - It just happens that China is leading the charge on this one.

The magic is NOT about the model. The magic is creating the guardrails and infrastructure (ecosystem) around the model, and supporting practical use cases the model excels at, and making them so good to use that no user wants to switch to something "almost as good", even if it’s a touch cheaper.

Net-net, governments cannot stop software at the border. Import/Export controls will NOT be effective at stopping Chinese firms from distilling US models, even if the USA is successful at starving China of inference, which also will NOT happen. https://futurism.com/artificial-intelligence/ai-execs-quaking-boots-chinese-models #AI #LLMs #Kimi-K3 #Kimi #FrontierModels #USA #China #ImportControls #ExportControls #Software #Distillation #Open-Weight-Models #AIPloicy #Protectionism #AIRegulation #DeepSeek #Software #OpenSource #Moonshot

The Moose has Left the Woods!

OR when your business model is crumbling before your eyes - go BEG the government to "protect" your product!

AI Executives are shitting their pants given the release of Moonshot AI’s Kimi K3. I seem to recall a similar situation re DeepSeek-1 in Jan. 2025.

All the freaking out about "winning and losing" and whining about "distillation" re China are the WRONG things to focus on.

It is a fallacy to think US Frontier Labs would be successful in locking down AI models and charging high prices for access in perpetuity. It is inevitable that open weight models would catch up, and put SERIOUS pressure on pricing.

Welcome to software - It just happens that China is leading the charge on this one.

The magic is NOT about the model. The magic is creating the guardrails and infrastructure (ecosystem) around the model, and supporting practical use cases the model excels at, and making them so good to use that no user wants to switch to something "almost as good", even if it’s a touch cheaper.

Net-net, governments cannot stop software at the border. Import/Export controls will NOT be effective at stopping Chinese firms from distilling US models, even if the USA is successful at starving China of inference, which also will NOT happen. https://futurism.com/artificial-intelligence/ai-execs-quaking-boots-chinese-models #AI #LLMs #Kimi-K3 #Kimi #FrontierModels #USA #China #ImportControls #ExportControls #Software #Distillation #Open-Weight-Models #AIPloicy #Protectionism #AIRegulation #DeepSeek #Software #OpenSource #Moonshot

Qwen3.8 will ship at 2.4 trillion parameters with open weights coming "soon," but Alibaba disclosed neither the activated-parameter count nor the license terms. Moonshot published benchmarks for its comparable model three days earlier. https://www.implicator.ai/alibaba-claims-qwen3-8-is-second-only-to-fable-5/ #OpenWeights #AITransparency #FrontierModels

The #Trump administration is asserting more #control over the rollout of future #AImodels, dictating which companies and entities can access the latest #frontiermodels. This shift in power from tech giants like #Anthropic and #OpenAI comes amid concerns about #nationalsecurity and the rapid advancement of #AI technology, particularly from Chinese startups. The administration’s actions aim to strengthen #AIsecurity while fostering innovation. https://www.cnbc.com/2026/07/17/white-house-ai-access-anthropic-openai.html?eicker.news #tech #media #news

The #Trump administration is asserting more #control over the rollout of future #AImodels, dictating which companies and entities can access the latest #frontiermodels. This shift in power from tech giants like #Anthropic and #OpenAI comes amid concerns about #nationalsecurity and the rapid advancement of #AI technology, particularly from Chinese startups. The administration’s actions aim to strengthen #AIsecurity while fostering innovation. https://www.cnbc.com/2026/07/17/white-house-ai-access-anthropic-openai.html?eicker.news #tech #media #news

While attention has been on #frontierAImodels, developers are increasingly using #openweight models, particularly from Chinese firms, for #productionAI. This shift is driven by #cost considerations and the desire for #customisation and #control over #AI capabilities. The rise of #openmodels raises questions about the future relevance of #frontiermodels and the potential risks associated with widespread access to powerful AI systems. https://techcrunch.com/2026/07/14/the-real-ai-race-may-no-longer-be-at-the-frontier-open-models-hugging-face/?eicker.news #tech #media #news

While attention has been on #frontierAImodels, developers are increasingly using #openweight models, particularly from Chinese firms, for #productionAI. This shift is driven by #cost considerations and the desire for #customisation and #control over #AI capabilities. The rise of #openmodels raises questions about the future relevance of #frontiermodels and the potential risks associated with widespread access to powerful AI systems. https://techcrunch.com/2026/07/14/the-real-ai-race-may-no-longer-be-at-the-frontier-open-models-hugging-face/?eicker.news #tech #media #news

“Now that organisations have been weaned off earlier 'all you can eat' #subscription plans and onto 'pay-as-you-go' metered #token consumption, they're all in various stages of sticker shock.

Several talks at the conference discussed managing token costs, such as AJ Fisher's exploration of 'diffusion' models. Analogous to the diffusers used to generate images, they generate text at lighting speed, making them cheaper to operate while also being less accurate than the pricey and slower “autoregressive” #FrontierModels.

Fisher's solution? Use a low-quality model and make it iterate on a problem (that new classic, the #RalphWiggumLoop) until it gets a satisfactory solution. This approach delivers the same result as a full-fat model, for anywhere from one half to one tenth the spend. #Google released its #DiffusionGemma model, which produces text at prodigious speed, just days after Fisher's talk, giving everyone the ability to try this approach.” — #MarkPesce

#AI / #ArtificialIntelligence / #developers / #software / #RalphWiggens / #Simpsons <https://theregister.com/columnists/2026/06/17/developers-build-the-best-tools-for-developers-and-are-now-defanging-the-ai-menace/5255316>

“Now that organisations have been weaned off earlier 'all you can eat' #subscription plans and onto 'pay-as-you-go' metered #token consumption, they're all in various stages of sticker shock.

Several talks at the conference discussed managing token costs, such as AJ Fisher's exploration of 'diffusion' models. Analogous to the diffusers used to generate images, they generate text at lighting speed, making them cheaper to operate while also being less accurate than the pricey and slower “autoregressive” #FrontierModels.

Fisher's solution? Use a low-quality model and make it iterate on a problem (that new classic, the #RalphWiggumLoop) until it gets a satisfactory solution. This approach delivers the same result as a full-fat model, for anywhere from one half to one tenth the spend. #Google released its #DiffusionGemma model, which produces text at prodigious speed, just days after Fisher's talk, giving everyone the ability to try this approach.” — #MarkPesce

#AI / #ArtificialIntelligence / #developers / #software / #RalphWiggens / #Simpsons <https://theregister.com/columnists/2026/06/17/developers-build-the-best-tools-for-developers-and-are-now-defanging-the-ai-menace/5255316>

The real prices of frontier models. Tokens * Price, right?

https://playcode.io/blog/real-price-of-frontier-models

Comments: https://news.ycombinator.com/item?id=48896800

#HackerNews #frontiermodels #pricing #tokens #AIinsights #machinelearning #dataanalysis

The real prices of frontier models. Tokens * Price, right?

https://playcode.io/blog/real-price-of-frontier-models

Comments: https://news.ycombinator.com/item?id=48896800

#HackerNews #frontiermodels #pricing #tokens #AIinsights #machinelearning #dataanalysis

China continues to attract top level AI talent from the U.S. ....

- OpenAI researcher now chief AI scientist at Tencent
- Google DeepMind researcher moves to Alibaba
- VP of Research at Google DeepMind moves to ByteDance
- Meta AI Researcher moves to Chinese startup Moonshot and creates Kimi AI model https://www.cnbc.com/2026/06/05/china-may-move-toward-us-path-on-ai-as-firms-poach-employees.html #AI #China #BrainDrain #AITalent #AGI #Tencent #Alibaba #ByteDance #SiliconValley #FrontierModels #FrontierAI

China continues to attract top level AI talent from the U.S. ....

- OpenAI researcher now chief AI scientist at Tencent
- Google DeepMind researcher moves to Alibaba
- VP of Research at Google DeepMind moves to ByteDance
- Meta AI Researcher moves to Chinese startup Moonshot and creates Kimi AI model https://www.cnbc.com/2026/06/05/china-may-move-toward-us-path-on-ai-as-firms-poach-employees.html #AI #China #BrainDrain #AITalent #AGI #Tencent #Alibaba #ByteDance #SiliconValley #FrontierModels #FrontierAI

The declining potential for agency with LLMs

It’s fascinating watching Ethan Mollick get gradually depilled as each successive generation of frontier models minimises the potential role for human agency:

Importantly, it was just limited in how much work I did relative to the model, it was also limited in how much control I had over how the model did things, why the model chose particular approaches, or even how in-depth its results would be. The details of the AI’s decision making are not shown to me, and the process would be too long to even be worth following. The map required the AI to make judgement calls about hundreds of little choices, and it just made them, without me understanding the choices or having a chance to weigh in. In many ways, it is miraculous (I can always ask for edits at the end) on the other, it turns AI into the ultimate black box.

https://www.oneusefulthing.org/p/what-it-feels-like-to-work-with-mythos

Last year I called this working with a wizard: you chant the spell and something happens. With Fable the spell has gotten powerful enough that I am no longer sure I am the wizard. I am closer to a patron. I describe what I want, I pay for it, and I judge the result. The conjuring happens somewhere I cannot watch, in hundreds of small choices I never get a vote on. The work has shifted from process to outcome. I no longer steer; I commission.

https://www.oneusefulthing.org/p/what-it-feels-like-to-work-with-mythos

Honestly I guess I’m going through a similar process, with the exception that I was never as utopian about the underlying capabilities of the technology. But the process he describes here, which Milan Sturmer and I talk about as reducing the burden of articulation, should be seen as the key shift taking place in the models as they diffuse. It’s just not possible to exercise sustained agency over the emerging models in the way it was over say 2025 era Claude and ChatGPT models. This captures something very significant I think:

A patron commissions a single artist. Fable is closer to a whole studio, where I am the client who signs off on the final work without ever setting foot on the floor.

AI clearly heading into the Trough of Disillusionment! (per Gartner's Hype Cycle)

This is a major warning flag for all Frontier LLM Vendors. Clients are now beginning to assess less expensive and open alternatives - can you say DeepSeek! https://www.tomshardware.com/tech-industry/artificial-intelligence/openai-ceo-sam-altman-admits-ai-token-costs-are-becoming-a-huge-issue-company-seeks-improved-value-as-overspending-becomes-a-meme

"Chinese models are 10x to 30x cheaper than U.S. models. We’re already seeing evidence of this: Chinese models went from about 1% of developer usage in 2024 to more than 60% in May, and 80% of U.S. AI startups are now using Chinese open-source AI models." https://www.profgmedia.com/p/is-ai-more-expensive-than-the-employees #AI #AIAdoption #OpenAI #Altman #AITokens #DeepSeek #AIInvestment #AIBuildout #LLMs #Capital #Investment #StartUps #FrontierModels #ProfGMedia #Budgets #Gartner #HypeCycle

AI clearly heading into the Trough of Disillusionment! (per Gartner's Hype Cycle)

This is a major warning flag for all Frontier LLM Vendors. Clients are now beginning to assess less expensive and open alternatives - can you say DeepSeek! https://www.tomshardware.com/tech-industry/artificial-intelligence/openai-ceo-sam-altman-admits-ai-token-costs-are-becoming-a-huge-issue-company-seeks-improved-value-as-overspending-becomes-a-meme

"Chinese models are 10x to 30x cheaper than U.S. models. We’re already seeing evidence of this: Chinese models went from about 1% of developer usage in 2024 to more than 60% in May, and 80% of U.S. AI startups are now using Chinese open-source AI models." https://www.profgmedia.com/p/is-ai-more-expensive-than-the-employees #AI #AIAdoption #OpenAI #Altman #AITokens #DeepSeek #AIInvestment #AIBuildout #LLMs #Capital #Investment #StartUps #FrontierModels #ProfGMedia #Budgets #Gartner #HypeCycle

OpenAI frontier models and Codex are now available on AWS

https://openai.com/index/openai-frontier-models-and-codex-are-now-available-on-aws/

#HackerNews #OpenAI #AWS #Codex #machinelearning #frontiermodels

China’s New AI Shocks The World: Hits Top 10 Globally Overnight
A sudden disruption on the global leaderboard. A newly deployed foundational model originating from China has breached the top ten tier of international benchmark evaluations, signaling a massive acceleration in eastern training efficiencies despite western hardware sanctions.
#GlobalTech #ChinaTech #AIResearch #MachineLearning #TechGeopolitics #FrontierModels
https://www.technology-news-channel.com/chinas-new-ai-shocks-the-world-hits-top-10-globally-overnight/