[{"data":1,"prerenderedAt":541},["ShallowReactive",2],{"\u002Fblog\u002Fchallenges-facing-ai-frontier-models-part-two":3,"blog-all-posts":466},{"_path":4,"_dir":5,"_draft":6,"_partial":6,"_locale":7,"title":8,"description":9,"titleLines":10,"descriptionLines":14,"bodyLeadTitle":17,"publishedAt":18,"updatedAt":19,"tags":20,"coverImage":24,"coverAlt":25,"heroLayout":26,"heroTitleSize":27,"featured":6,"draft":6,"body":28,"_type":460,"_id":461,"_source":462,"_file":463,"_stem":464,"_extension":465},"\u002Fblog\u002Fchallenges-facing-ai-frontier-models-part-two","blog",false,"","Challenges Facing AI Frontier Models Part Two","Frontier AI models must overcome constraints in context, agent design, data quality, energy, and open-source competition.",[11,12,13],"Challenges Facing","AI Frontier Models","Part Two",[15,16],"Frontier AI models face multiple","technical and competitive challenges","The Next Frontier","2026-06-25","2026-06-26",[21,22,23],"artificial intelligence","frontier models","AI safety","\u002Fblog\u002Fchallenges-facing-ai-frontier-models-part-two\u002Fcover.jpg","A humanoid robot ascending a snowy mountain ridge toward the sunset","overlay","compact",{"type":29,"children":30,"toc":452},"root",[31,52,59,71,116,126,136,141,147,152,169,181,186,194,200,205,210,215,258,263,275,292,298,303,311,316,351,356,366,372,377,382,417,427,437,442],{"type":32,"tag":33,"props":34,"children":35},"element","p",{},[36,39,45,47],{"type":37,"value":38},"text","Building an intelligent AI model is only the beginning. Scaling it into a ",{"type":32,"tag":40,"props":41,"children":42},"strong",{},[43],{"type":37,"value":44},"reliable, trustworthy, and economically sustainable platform",{"type":37,"value":46}," presents an entirely different set of challenges. Every frontier AI company faces difficult tradeoffs. Solving these problems will require breakthroughs not only in artificial intelligence, but also in computer science, engineering, and energy infrastructure. ",{"type":32,"tag":40,"props":48,"children":49},{},[50],{"type":37,"value":51},"The companies that overcome these challenges will define the next era of AI.",{"type":32,"tag":53,"props":54,"children":56},"h2",{"id":55},"context-window",[57],{"type":37,"value":58},"Context Window",{"type":32,"tag":33,"props":60,"children":61},{},[62,64,69],{"type":37,"value":63},"The context window is the amount of information an AI model can consider at one time while generating a response. ",{"type":32,"tag":40,"props":65,"children":66},{},[67],{"type":37,"value":68},"It is the model's working memory for the current conversation or task.",{"type":37,"value":70}," The context window can include:",{"type":32,"tag":72,"props":73,"children":74},"ul",{},[75,84,92,100,108],{"type":32,"tag":76,"props":77,"children":78},"li",{},[79],{"type":32,"tag":40,"props":80,"children":81},{},[82],{"type":37,"value":83},"The current prompt",{"type":32,"tag":76,"props":85,"children":86},{},[87],{"type":32,"tag":40,"props":88,"children":89},{},[90],{"type":37,"value":91},"Previous messages in the conversation",{"type":32,"tag":76,"props":93,"children":94},{},[95],{"type":32,"tag":40,"props":96,"children":97},{},[98],{"type":37,"value":99},"Uploaded documents or files",{"type":32,"tag":76,"props":101,"children":102},{},[103],{"type":32,"tag":40,"props":104,"children":105},{},[106],{"type":37,"value":107},"The AI's previous responses",{"type":32,"tag":76,"props":109,"children":110},{},[111],{"type":32,"tag":40,"props":112,"children":113},{},[114],{"type":37,"value":115},"Instructions provided by an application or developer",{"type":32,"tag":33,"props":117,"children":118},{},[119,121],{"type":37,"value":120},"Every frontier and open-source AI model has a finite context window that limits the amount of information it can actively attend to while generating a response. Information outside that window is no longer directly available unless it is retrieved or summarized. Although some frontier models now support context windows approaching or exceeding one million tokens, larger contexts do not eliminate the underlying challenge. ",{"type":32,"tag":40,"props":122,"children":123},{},[124],{"type":37,"value":125},"As conversations become longer, models must distribute attention across vastly more information, making retrieval, prioritization, and multi-step reasoning increasingly difficult.",{"type":32,"tag":33,"props":127,"children":128},{},[129,131],{"type":37,"value":130},"Research has shown that models often perform worse when important information is buried in the middle of very long contexts, a phenomenon commonly referred to as ",{"type":32,"tag":40,"props":132,"children":133},{},[134],{"type":37,"value":135},"\"lost in the middle.\"",{"type":32,"tag":33,"props":137,"children":138},{},[139],{"type":37,"value":140},"As companies and users adopt AI models, the need for increasingly large context windows becomes apparent. Size limitations make long-form planning, software engineering, legal analysis, and large-document review particularly challenging.",{"type":32,"tag":53,"props":142,"children":144},{"id":143},"agentification-of-tasks-processes",[145],{"type":37,"value":146},"Agentification of Tasks & Processes",{"type":32,"tag":33,"props":148,"children":149},{},[150],{"type":37,"value":151},"To agentify a process or task means transforming it from one executed by a human into a structured workflow in which an AI agent can reason, plan, decide, and act toward achieving a goal.",{"type":32,"tag":33,"props":153,"children":154},{},[155,157,162,164],{"type":37,"value":156},"Many organizations have rushed into AI initiatives believing that business processes can simply be ",{"type":32,"tag":40,"props":158,"children":159},{},[160],{"type":37,"value":161},"\"wrapped\"",{"type":37,"value":163}," with a large language model to create an intelligent agent. While it is easy to use AI to quickly spin up an application that performs one process or task, ",{"type":32,"tag":40,"props":165,"children":166},{},[167],{"type":37,"value":168},"developing a more complex AI agent with long-form reasoning, memory, and advanced logic is quite difficult.",{"type":32,"tag":33,"props":170,"children":171},{},[172,174,179],{"type":37,"value":173},"Many businesses will state that they have developed an AI agent, ",{"type":32,"tag":40,"props":175,"children":176},{},[177],{"type":37,"value":178},"but in reality they have only created an automated process.",{"type":37,"value":180}," Developing an agent to perform a simple task is one thing; trusting it to make multiple decisions across a sequence of operations without human input is quite another.",{"type":32,"tag":33,"props":182,"children":183},{},[184],{"type":37,"value":185},"A workflow that follows predefined steps is automation. An AI agent must perceive, reason, plan, adapt to changing conditions, recover from failures, and determine the next action without explicit human direction.",{"type":32,"tag":33,"props":187,"children":188},{},[189],{"type":32,"tag":40,"props":190,"children":191},{},[192],{"type":37,"value":193},"True agentification, designing \"human-level\" logic and reasoning into an autonomous agent, is extremely challenging and requires considerable investment in engineering, time, and money.",{"type":32,"tag":53,"props":195,"children":197},{"id":196},"ai-slop",[198],{"type":37,"value":199},"AI Slop",{"type":32,"tag":33,"props":201,"children":202},{},[203],{"type":37,"value":204},"AI slop is an informal term used to describe low-quality, mass-produced content created with AI and published with little or no meaningful human review or originality. This applies to text, images, videos, music, websites, and social media posts that prioritize quantity and engagement over accuracy, creativity, or usefulness.",{"type":32,"tag":33,"props":206,"children":207},{},[208],{"type":37,"value":209},"Not all AI-generated content is \"AI slop.\" AI is extremely useful for the research, analysis, and creation of content when paired with proper human oversight. It helps people efficiently create high-quality work.",{"type":32,"tag":33,"props":211,"children":212},{},[213],{"type":37,"value":214},"AI slop becomes challenging when it displays the following traits:",{"type":32,"tag":72,"props":216,"children":217},{},[218,226,234,242,250],{"type":32,"tag":76,"props":219,"children":220},{},[221],{"type":32,"tag":40,"props":222,"children":223},{},[224],{"type":37,"value":225},"Generic information",{"type":32,"tag":76,"props":227,"children":228},{},[229],{"type":32,"tag":40,"props":230,"children":231},{},[232],{"type":37,"value":233},"Repetitive content",{"type":32,"tag":76,"props":235,"children":236},{},[237],{"type":32,"tag":40,"props":238,"children":239},{},[240],{"type":37,"value":241},"Low originality",{"type":32,"tag":76,"props":243,"children":244},{},[245],{"type":32,"tag":40,"props":246,"children":247},{},[248],{"type":37,"value":249},"Factual errors",{"type":32,"tag":76,"props":251,"children":252},{},[253],{"type":32,"tag":40,"props":254,"children":255},{},[256],{"type":37,"value":257},"Fabricated evidence",{"type":32,"tag":33,"props":259,"children":260},{},[261],{"type":37,"value":262},"The proliferation of AI slop can increase misinformation and content overload while reducing trust.",{"type":32,"tag":33,"props":264,"children":265},{},[266,268,273],{"type":37,"value":267},"The greatest risk may be to the frontier models themselves, which interact with data from social media and the Internet as a whole. ",{"type":32,"tag":40,"props":269,"children":270},{},[271],{"type":37,"value":272},"Over time, frontier models can become poisoned by the very AI slop their users create.",{"type":37,"value":274}," As new generations of models are trained on increasingly recent Internet content, a growing proportion of that data contains AI-generated material. If it is not carefully filtered, future training datasets may contain synthetic content produced by previous generations of AI models.",{"type":32,"tag":33,"props":276,"children":277},{},[278,280,285,287],{"type":37,"value":279},"Researchers increasingly refer to this as ",{"type":32,"tag":40,"props":281,"children":282},{},[283],{"type":37,"value":284},"model collapse or data contamination",{"type":37,"value":286},", in which future models learn from synthetic rather than human-generated data. Without careful curation, this feedback loop can ",{"type":32,"tag":40,"props":288,"children":289},{},[290],{"type":37,"value":291},"gradually reduce originality, diversity, and factual quality.",{"type":32,"tag":53,"props":293,"children":295},{"id":294},"energy",[296],{"type":37,"value":297},"Energy",{"type":32,"tag":33,"props":299,"children":300},{},[301],{"type":37,"value":302},"AI inference represents a significant and recurring expenditure for frontier-model hyperscalers. From GPUs and accelerators to complex data centers, every element requires considerable capital investment. While training frontier models requires enormous bursts of computational power, inference represents the long-term economic challenge because every user request consumes compute resources.",{"type":32,"tag":33,"props":304,"children":305},{},[306],{"type":32,"tag":40,"props":307,"children":308},{},[309],{"type":37,"value":310},"The greatest expense, along with the greatest potential scarcity, is energy. AI will likely become one of the largest consumers of electricity over the next decade.",{"type":32,"tag":33,"props":312,"children":313},{},[314],{"type":37,"value":315},"Many projections estimate that AI could require hundreds of gigawatts of installed data-center capacity globally by the 2030s if demand continues growing rapidly. Electricity is needed not only for computing, but also for:",{"type":32,"tag":72,"props":317,"children":318},{},[319,327,335,343],{"type":32,"tag":76,"props":320,"children":321},{},[322],{"type":32,"tag":40,"props":323,"children":324},{},[325],{"type":37,"value":326},"Cooling",{"type":32,"tag":76,"props":328,"children":329},{},[330],{"type":32,"tag":40,"props":331,"children":332},{},[333],{"type":37,"value":334},"Networking",{"type":32,"tag":76,"props":336,"children":337},{},[338],{"type":32,"tag":40,"props":339,"children":340},{},[341],{"type":37,"value":342},"Power conversion",{"type":32,"tag":76,"props":344,"children":345},{},[346],{"type":32,"tag":40,"props":347,"children":348},{},[349],{"type":37,"value":350},"Security systems",{"type":32,"tag":33,"props":352,"children":353},{},[354],{"type":37,"value":355},"The many uses of multimodal AI in video generation and gaming are especially energy intensive. Generating HD video, 3D worlds, realistic simulations, and engineering designs is computationally expensive.",{"type":32,"tag":33,"props":357,"children":358},{},[359,361],{"type":37,"value":360},"As AI applications become richer, the energy required per request may increase unless hardware improvements offset it. ",{"type":32,"tag":40,"props":362,"children":363},{},[364],{"type":37,"value":365},"Going forward, one of the greatest challenges facing AI will be power cost and availability.",{"type":32,"tag":53,"props":367,"children":369},{"id":368},"open-source-models",[370],{"type":37,"value":371},"Open-Source Models",{"type":32,"tag":33,"props":373,"children":374},{},[375],{"type":37,"value":376},"Open-source models present one of the most significant long-term competitive challenges facing frontier-model providers. If established businesses and startups can use free open models that achieve near parity in the quality of their inference, the choice becomes obvious.",{"type":32,"tag":33,"props":378,"children":379},{},[380],{"type":37,"value":381},"Software economics often favor open ecosystems:",{"type":32,"tag":72,"props":383,"children":384},{},[385,393,401,409],{"type":32,"tag":76,"props":386,"children":387},{},[388],{"type":32,"tag":40,"props":389,"children":390},{},[391],{"type":37,"value":392},"Linux reduced the dominance of proprietary Unix.",{"type":32,"tag":76,"props":394,"children":395},{},[396],{"type":32,"tag":40,"props":397,"children":398},{},[399],{"type":37,"value":400},"Apache became a standard web server over IIS.",{"type":32,"tag":76,"props":402,"children":403},{},[404],{"type":32,"tag":40,"props":405,"children":406},{},[407],{"type":37,"value":408},"PostgreSQL competes effectively with commercial databases.",{"type":32,"tag":76,"props":410,"children":411},{},[412],{"type":32,"tag":40,"props":413,"children":414},{},[415],{"type":37,"value":416},"Kubernetes became the standard orchestration platform.",{"type":32,"tag":33,"props":418,"children":419},{},[420,422],{"type":37,"value":421},"Open models benefit from thousands of researchers and engineers collectively improving the model itself. Innovation becomes distributed rather than centralized within a single company. ",{"type":32,"tag":40,"props":423,"children":424},{},[425],{"type":37,"value":426},"Proprietary advantages become narrower if an open model achieves, for example, 95% of the performance at a fraction of the cost; many use cases will favor the less expensive option.",{"type":32,"tag":33,"props":428,"children":429},{},[430,435],{"type":32,"tag":40,"props":431,"children":432},{},[433],{"type":37,"value":434},"Open-source models put pressure on the profitability of proprietary frontier models.",{"type":37,"value":436}," Whether they pose an existential economic threat depends on whether frontier developers can continue to maintain meaningful capability advantages.",{"type":32,"tag":33,"props":438,"children":439},{},[440],{"type":37,"value":441},"Training state-of-the-art models remains extraordinarily expensive regardless of whether the resulting weights are openly released. The long-term sustainability of releasing increasingly capable models under permissive licenses remains an open economic question.",{"type":32,"tag":33,"props":443,"children":444},{},[445,447],{"type":37,"value":446},"Training and inference costs also remain high for companies that release open-source models. The question remains: ",{"type":32,"tag":40,"props":448,"children":449},{},[450],{"type":37,"value":451},"how long can open-source companies afford to release their models into the wild entirely for free?",{"title":7,"searchDepth":453,"depth":453,"links":454},2,[455,456,457,458,459],{"id":55,"depth":453,"text":58},{"id":143,"depth":453,"text":146},{"id":196,"depth":453,"text":199},{"id":294,"depth":453,"text":297},{"id":368,"depth":453,"text":371},"markdown","content:blog:challenges-facing-ai-frontier-models-part-two.md","content","blog\u002Fchallenges-facing-ai-frontier-models-part-two.md","blog\u002Fchallenges-facing-ai-frontier-models-part-two","md",[467,479,488,490,500,508,520,531],{"_path":468,"title":469,"description":470,"cardTitle":471,"publishedAt":472,"tags":473,"coverImage":477,"coverAlt":478,"featured":6},"\u002Fblog\u002Funderstanding-your-true-attack-surface-part-two","Understanding Your True Attack Surface: Part Two","Few companies understand how they look to an attacker. Your Internet attack surface is exactly what attackers see first.","Understanding Your True Attack Surface Part Two","2026-07-28",[474,475,476],"attack surface","internet security","cybersecurity strategy","\u002Fblog\u002Funderstanding-your-true-attack-surface-part-two\u002Fcover.jpg","Cyber operations control tower monitoring a connected city and its Internet-facing attack surface",{"_path":480,"title":481,"description":482,"publishedAt":19,"tags":483,"coverImage":486,"coverAlt":487,"featured":6},"\u002Fblog\u002Ftop-cyber-threats-facing-businesses-in-2026","Top Cyber Threats Facing Businesses in 2026","Businesses face AI-assisted hacking, interception-proxy phishing, dark web breach data, and API attacks in 2026.",[484,485,476],"cyber threats","business security","\u002Fblog\u002Ftop-cyber-threats-facing-businesses-in-2026\u002Fcover.jpg","A city skyline beneath a massive storm cloud embedded with cybersecurity symbols and lightning",{"_path":4,"title":8,"description":9,"publishedAt":18,"tags":489,"coverImage":24,"coverAlt":25,"featured":6},[21,22,23],{"_path":491,"title":492,"description":493,"cardTitle":494,"publishedAt":495,"tags":496,"coverImage":498,"coverAlt":499,"featured":6},"\u002Fblog\u002Funderstanding-your-true-attack-surface-part-one","Understanding Your True Attack Surface: Part One","The modern attack surface extends far beyond the corporate firewall.","Understanding Your True Attack Surface Part One","2026-06-21",[474,497,476],"network security","\u002Fblog\u002Funderstanding-your-true-attack-surface-part-one\u002Fcover.jpg","Security leader reviewing an enterprise attack surface dashboard spanning cloud, applications, remote offices, devices, patching, identities, and internet exposure",{"_path":501,"title":502,"description":503,"publishedAt":504,"tags":505,"coverImage":506,"coverAlt":507,"featured":6},"\u002Fblog\u002Fchallenges-facing-ai-frontier-models","Challenges Facing AI Frontier Models Part One","Frontier AI models face a series of challenges that threaten their viability","2026-06-12",[21,22,23],"\u002Fblog\u002Fchallenges-facing-ai-frontier-models\u002Fcover.jpg","A humanoid robot standing among clouds beneath a blue sky",{"_path":509,"title":510,"description":511,"cardTitle":512,"publishedAt":513,"tags":514,"coverImage":518,"coverAlt":519,"featured":6},"\u002Fblog\u002Fexecutive-operational-security-opsec-part-one-setting-the-foundation","Executive Operational Security Part One: Setting the Foundation","Executive OPSEC protects senior leaders, privileged access, sensitive information, communications, travel, and personal activity from targeted threats.","Executive Operational Security Part One","2026-06-10",[515,516,517],"executive security","OPSEC","risk management","\u002Fblog\u002Fexecutive-operational-security-opsec-part-one-setting-the-foundation\u002Fcover.jpg","Executive reviewing his phone beside a laptop in a private office",{"_path":521,"title":522,"description":523,"publishedAt":524,"tags":525,"coverImage":529,"coverAlt":530,"featured":6},"\u002Fblog\u002Fevaluating-identity-before-attackers-do","Evaluating Identity Before Attackers Do","An IAM assessment evaluates authentication, access controls, user behavior, and identity resilience before attackers can exploit them.","2026-06-07",[526,527,528],"identity security","IAM assessment","access management","\u002Fblog\u002Fevaluating-identity-before-attackers-do\u002Fcover.jpg","Open cyber vault displaying identity security controls and authentication monitoring",{"_path":532,"title":533,"description":534,"publishedAt":535,"tags":536,"coverImage":539,"coverAlt":540,"featured":6},"\u002Fblog\u002Fthe-necessity-of-an-enterprise-risk-assessment","The Necessity of an Enterprise Risk Assessment","Effective Cybersecurity is both a Technical and Procedural Endeavor","2026-05-26",[537,517,538],"enterprise risk assessment","cybersecurity governance","\u002Fblog\u002Fthe-necessity-of-an-enterprise-risk-assessment\u002Fcover.jpg","CISO in an aircraft cockpit overlooking clouds with technical and procedural cybersecurity displays",1785277054252]