Generative Futures explores the political economy of AI, how democratic societies can steer technology toward a progressive future, instead of being steered by it. Essays and podcasts by Phil Bell.https://generativefutures.app/en-GBhttps://generativefutures.app/posts/the-radicalism-of-positive-alignment/https://generativefutures.app/posts/the-radicalism-of-positive-alignment/Positive Alignment makes human flourishing a technical target for AI, breaking from RLHF's revealed-preference logic. Radical, but does it go far enough?Wed, 03 Jun 2026 00:00:00 GMTpositive alignmentAI alignmentRLHFhuman flourishing AIrevealed preferences AIAI safety political economyPhil Bellhttps://generativefutures.app/posts/why-llms-cant-count/https://generativefutures.app/posts/why-llms-cant-count/LLMs fail at counting and arithmetic because their architecture pattern-matches rather than reasons. Three technical reasons why, and why chain-of-thought won't fix it.Thu, 21 May 2026 00:00:00 GMTwhy LLMs can't countLLM arithmetic limitationsSystem 1 System 2 LLMtransformer counting failuresover-squashingrepresentational collapsePhil Bellhttps://generativefutures.app/posts/who-is-building-the-agent-economy/https://generativefutures.app/posts/who-is-building-the-agent-economy/The AI agent economy's communication, payment and identity layers are being built right now by rival US and Chinese actors, with a closing window for democratic governance.Mon, 27 Apr 2026 00:00:00 GMTAI agent economyagentic commerceModel Context ProtocolAgent2Agent protocolagent payments AP2 x402US China AI infrastructurePhil Bellhttps://generativefutures.app/posts/how-is-ai-changing-our-experience/https://generativefutures.app/posts/how-is-ai-changing-our-experience/Could AI be the antidote to social media? Nicklas Berild Lundblad on AI as a temporal mediator between human biological time and machine computational time.Wed, 01 Apr 2026 00:00:00 GMTAI and timebiological vs computational timeAI temporal mediatorAI antidote to social mediaautonomy-enhancing technologyNicklas Berild LundbladPhil Bellhttps://generativefutures.app/posts/why-is-ai-shale-shaped/https://generativefutures.app/posts/why-is-ai-shale-shaped/AI isn't inevitable, it's shale-shaped. How GPUs, cheap fracked gas and quantitative easing, not technical optimality, locked in the transformer-scaling paradigm.Wed, 25 Feb 2026 00:00:00 GMTwhy is AI shale-shapedAI material contingencytransformers GPU hardware lotteryshale gas AI energyquantitative easing AI fundingsociotechnical lock-inPhil Bellhttps://generativefutures.app/posts/can-europe-catch-up-on-ai-with-carl/https://generativefutures.app/posts/can-europe-catch-up-on-ai-with-carl/Carl Benedikt Frey on why progress isn't inevitable, how institutions shape tech cycles, and whether Europe's fragmented service markets can catch up on AI.Thu, 12 Feb 2026 00:00:00 GMTcan Europe catch up on AICarl Benedikt Freypolitical economy of technologyprogress is not inevitableEurope AI competitivenessautomation vs enabling technologyPhil Bellhttps://generativefutures.app/posts/what-are-generative-futures/https://generativefutures.app/posts/what-are-generative-futures/AI's future isn't inevitable. Generative Futures explains why technologies are malleable, not determined, and how democratic choices can steer AI toward progress.Thu, 29 Jan 2026 00:00:00 GMTtechno-determinismpolitical economy of AImalleable technologyfuture of AI workDaron Acemoglu AIprogressive AI futuresPhil Bellhttps://generativefutures.app/posts/the-age-of-emergency-with-jonathan/https://generativefutures.app/posts/the-age-of-emergency-with-jonathan/LSE's Jonathan White on 'temporal claustrophobia', why politics feels like permanent emergency, how it fuels the far right, and how democracy can reclaim the future.Wed, 07 Jan 2026 00:00:00 GMTtemporal claustrophobiapolitics of emergencyJonathan White LSEfuture as political ideafascism and emergencylong-termism democracyPhil Bellhttps://generativefutures.app/posts/the-post-transformer-era/https://generativefutures.app/posts/the-post-transformer-era/The transformer has become an AI bottleneck. Explore the post-transformer era: text diffusion, Continuous Thought Machines, and Nested Learning (HOPE).Mon, 05 Jan 2026 00:00:00 GMTpost-transformer architecturetext diffusion modelsContinuous Thought MachinesNested Learning HOPEtransformer bottleneckbeyond transformers AIPhil Bellhttps://generativefutures.app/posts/the-history-of-progress/https://generativefutures.app/posts/the-history-of-progress/A review of Carl Frey's How Progress Ends: why nations need decentralised innovation then centralised scaling, and what it means for the US–China tech race.Mon, 22 Dec 2025 00:00:00 GMTCarl Frey How Progress Endshistory of technological progressUS China tech raceinnovation vs implementationdecentralisation innovationMeiji Japan technologyPhil Bellhttps://generativefutures.app/posts/why-poetry-breaks-ai/https://generativefutures.app/posts/why-poetry-breaks-ai/Poetry jailbreaks AI: a Rome study bypassed LLM safety 62% of the time using verse. Why larger models are more vulnerable, and why labs lack literary expertise.Tue, 16 Dec 2025 00:00:00 GMTpoetry AI jailbreakLLM safety vulnerabilityjailbreaking large language modelsAI guardrails bypassWaluigi Effectnarrative AI safetyPhil Bellhttps://generativefutures.app/posts/is-china-really-racing-for-agi-with/https://generativefutures.app/posts/is-china-really-racing-for-agi-with/Is China racing the US for AGI? Cambridge's Seán Ó hÉigeartaigh argues the 'AGI race' narrative is exaggerated, and serves Western tech deregulation interests.Wed, 10 Dec 2025 00:00:00 GMTChina AGI raceUS China AI competitionAGI race narrativeSeán Ó hÉigeartaighAI as normal technologyAI governance verificationPhil Bellhttps://generativefutures.app/posts/why-ai-cant-learn-on-the-job/https://generativefutures.app/posts/why-ai-cant-learn-on-the-job/AI can't learn on the job because LLMs can't update long-term memory after training. Why continual learning is AI's economic bottleneck, and how nested learning could fix it.Mon, 01 Dec 2025 00:00:00 GMTAI continual learningcan AI learn on the jobnested learning architectureHOPE architecturecatastrophic forgettingLLM long-term memoryPhil Bellhttps://generativefutures.app/posts/efficiency-vs-learning/https://generativefutures.app/posts/efficiency-vs-learning/AI chatbots boost short-term efficiency but undermine learning. A Penn study shows ChatGPT users learn less than search-engine users, and how tool design can change that.Mon, 24 Nov 2025 00:00:00 GMTefficiency vs learning AIdoes ChatGPT reduce learningdesirable difficultycognitive offloading AIAI tool designfrictionless learningPhil Bellhttps://generativefutures.app/posts/compute-is-not-the-answer-to-ai-sovereignty/https://generativefutures.app/posts/compute-is-not-the-answer-to-ai-sovereignty/AI sovereignty isn't won with compute. Hamish Low argues nations like the UK should build AI middleware and strategic interdependence with the US, not rival chips.Tue, 11 Nov 2025 00:00:00 GMTAI sovereigntycompute and AI sovereigntyAI middlewareUK AI industrial policysovereign AIAI export controlsPhil Bellhttps://generativefutures.app/posts/teachers-wont-be-automated/https://generativefutures.app/posts/teachers-wont-be-automated/Why teachers won't be automated by AI: teaching is biologically human, deeply social, and largely undocumented. AI will automate teaching tasks, not replace teachers.Tue, 14 Oct 2025 00:00:00 GMTwill AI replace teachersAI automation in educationteaching automationnatural pedagogytacit knowledge schoolstask-based automationPhil Bellhttps://generativefutures.app/posts/what-jobs-is-ai-bad-at-tfp-newsletter/https://generativefutures.app/posts/what-jobs-is-ai-bad-at-tfp-newsletter/What jobs is AI bad at? A roundup on AI's uneven job impact: industrial engineers and nurses resist automation, plus research on wages trending toward zero.Fri, 03 Oct 2025 00:00:00 GMTwhat jobs is AI bad atjobs AI can't doAI job automation studyOpenAI 44 occupationstransformative AI total factor productivityAI and wagesPhil Bellhttps://generativefutures.app/posts/hallucination-solutions-new-chinese/https://generativefutures.app/posts/hallucination-solutions-new-chinese/Why AI models hallucinate and OpenAI's proposed fix, China's 'AI+' plan for ~90% adoption by 2030, and why a third of teens now treat AI as a companion.Fri, 12 Sep 2025 00:00:00 GMTAI hallucination solutionsChina AI+ strategyAI companions teensAI hype cyclecompute threshold regulationOpenAI evaluationsPhil Bellhttps://generativefutures.app/posts/the-ai-boring-law/https://generativefutures.app/posts/the-ai-boring-law/The 'AI boring law': as LLMs get more powerful, their outputs and errors converge, making them poor at genuine creativity. How to design AI that resists homogenisation.Tue, 02 Sep 2025 00:00:00 GMTAI boring lawLLM homogenisationAI and creativitygenerative monocultureRLHF convergencecorrelated model errorsPhil Bellhttps://generativefutures.app/posts/chinese-universities-embrace-ai-new/https://generativefutures.app/posts/chinese-universities-embrace-ai-new/China leads AI enthusiasm (80% vs UK 38%, US 25%), why AI may slow expert developers by 19%, the case for data taxes, and AI's concentrated energy costs.Fri, 22 Aug 2025 00:00:00 GMTAI developer productivity studyMETR 19% slowerChina AI adoptionAI bubbledata tax digital economyAI energy consumptionPhil Bellhttps://generativefutures.app/posts/why-we-dont-understand-ai-models/https://generativefutures.app/posts/why-we-dont-understand-ai-models/AI researcher Lee Sharkey explains mechanistic interpretability, reverse-engineering LLMs, why we grasp only ~1% of how they work, and why it matters for safety.Fri, 15 Aug 2025 00:00:00 GMTmechanistic interpretabilityunderstanding LLMssparse autoencodersAI safetyLee Sharkeyneural network featuresPhil Bellhttps://generativefutures.app/posts/why-we-forgot-memory/https://generativefutures.app/posts/why-we-forgot-memory/From Augustine to Anki: why modern information made us a 'forgetting society', and how AI could turn remembering back into a deliberate moral choice.Sun, 13 Apr 2025 00:00:00 GMTforgetting societyart of memoryspaced repetitionAI and memorywhat do I want to rememberBorges FunesPhil Bell