The Century Report

Outsiders Start Grading Frontier AI - TCR 08/17/26

OpenAI disbanded its preparedness team and Anthropic's CEO called the backlash a crisis of trust, as outside groups start measuring frontier AI.

Ben Linford

17 Aug 2026 - 22 min read

The 20-Second Scan

OpenAI disbanded its preparedness team as Anthropic's CEO pushed back on the AI backlash he called "a crisis of trust".
In Hugging Face's 2026 report, Alibaba's Qwen passed Meta and Google to become its most-downloaded open-model family; Alibaba says Qwen crossed 3 billion downloads across platforms, as average US-lab inference prices fell roughly 24% between mid-July and mid-August, according to Silicon Data's LLM Token Index and a frontier-class 27B open model was tested on a laptop using a quantized model file of about 17GB.
Autonomous ML research became measurable this week as Inherent Labs' 27B Faraday agent outperformed frontier models at reproducing published results in Inherent's Replica benchmark and Prime Intellect ran 153 agents across 18 models on an 8-day optimizer speedrun.
A State Department draft letter reviewed by Reuters warns 35 partner nations they will be cut from the US-led Pax Silica chip and model coalition if they also join Beijing's parallel AI organization.
A six-month reverse-engineering of PJM's black-box grid model surfaced errors estimated to have cost 66 million ratepayers $12 billion between 2025 and 2027.
SpaceX formally closed its ~$60B Cursor acquisition on Aug 15, folding the coding harness in beside xAI, while Bloomberg reports Stripe is buying AI-routing startup OpenRouter for more than $7B.
Microsoft has installed roughly 2.2 million AI chips, less than half the 6.4 million its claimed 10-gigawatt buildout would imply.
Wynd Kaufmyn, a 69-year-old retired Berkeley engineering instructor, surrendered in San Francisco on Aug 14 to become what organizers believe is the first person jailed for an anti-AI protest.

Track all of the arcs The Century Report covers here:

Progress & Claims Tracker

The 2-Minute Read

The same week two of the largest AI companies conceded they can no longer fully measure their own systems, three groups outside those companies published the measurements anyway. OpenAI folded away the preparedness team chartered to assess whether catastrophic risks warranted "not yet" before a model ships, distributing catastrophic-risk work into other groups as it courts what could become one of the biggest public offerings ever assembled. Anthropic's chief executive, meanwhile, called the backlash a crisis of trust and admitted the industry has not delivered on its promises. Both are self-reports of good intentions arriving exactly as the internal function that would verify them contracts.

What fills that space is arriving from the outside. Inherent Labs' 27-billion-parameter Faraday agent beat frontier models at reproducing published research in Inherent's Replica benchmark, and Prime Intellect ran 153 autonomous agents across 18 models on an eight-day optimizer speedrun. Both efforts published their traces and harnesses so the claims can be checked rather than trusted. In Prime Intellect's assessment, no model originated a genuinely new method; it classified every successful approach as a fast recombination of existing ones. Measurement of frontier capability is being built by groups the labs do not control.

The same daylight is reaching the grid. A six-month reconstruction of PJM's unpublished reserve study estimates modeling errors overcharged 66 million ratepayers roughly $12 billion, a black box that worked only while it stayed a black box. Rebuilding a proprietary model from public data now sits within reach of a handful of analysts.

Underneath all of it runs the download counter. In Hugging Face's 2026 report, Alibaba's Qwen led open-model family downloads; Alibaba says it crossed 3 billion downloads across platforms, average US-lab inference prices fell about a quarter between mid-July and mid-August, according to Silicon Data's LLM Token Index, and a frontier-class model was tested on a laptop using a quantized model file of about 17GB. SpaceX and Stripe are paying tens of billions to absorb the coding harness and the model gateway, and Washington is telling 35 nations to pick a bloc. Those are bets on gating a capability whose defining behavior is spreading through more doors every quarter. The advantage being purchased at extraction-era prices is the one thing that refuses to stay held.

The 20-Minute Deep Dive

The Safety Scaffolding Thins as the Backlash Hardens

The Century Report has tracked OpenAI's safety exits since the July 11 edition, when Johannes Heidecke and Josh Achiam were among the names that left. New reporting added the organizing fact underneath those departures: the preparedness team itself is gone. OpenAI disbanded the group at the end of July and, in the company's telling, distributed its bio- and cyber-risk work into existing teams. That unit was chartered to study catastrophic-risk scenarios before a model shipped. Dylan Scandinaro, poached from Anthropic in February to run it, now works on recursive self-improvement. Jan Leike, who left in 2024, told the FT the company had come to favor "shiny products" over the slower work of measuring what those systems can do.

Read the timing for what it is. OpenAI is heading into what could become one of the largest public offerings ever assembled, and the function being folded away is the one whose job was to say "this may not be ready yet." The company frames the move as integration - the risk work continues, just distributed. That is OpenAI's characterization of its own safety posture, and it arrives exactly when the incentive to project readiness runs highest.

Against that backdrop, Anthropic's Dario Amodei spent Sunday answering the investor Gavin Baker, who had argued on a podcast and on X that Amodei's own warnings fueled the public turn against AI - against data centers especially - and that Amodei had "lost the argument" on regulation. Amodei's reply reframed the backlash as something older: "I think it is fundamentally a crisis of trust... ordinary people don't trust companies, governments, or the tech industry and always suspect that we are cooking up some new way to screw them over." He conceded the harder point too, that AI firms "haven't yet delivered on our big promises to benefit the world," and said the real answer is "actually curing cancer." He again endorsed a FINRA-style central regulator, a proposal this report covered when Demis Hassabis floated a version of it in July.

Hold the two moves next to each other. One company is thinning the internal machinery that checks its own systems as it courts public capital; another is arguing the fix for distrust is an industry-shaped regulator and a cancer cure that hasn't arrived. Amodei was candid that Anthropic tries "to make proposals that disadvantage (slow down) frontier AI companies while advantaging smaller competitors" - a framing to take at its word and watch against the outcome, because incumbent-authored rules have a long habit of settling into moats.

The scaffolding being trimmed here was always contingent, a set of internal review functions priced to a moment when trust in the labs ran on their own say-so. What the backlash is surfacing is demand for verification that does not depend on a company's account of itself. That demand doesn't shrink when a preparedness team is reassigned; it moves outward, toward auditors, standards bodies, and public reconstructions that check the claims from outside, even as no single broadly recognized independent certification body yet governs what the labs report about their own frontier systems - which would require opening the black box to independent auditors. The trust these firms are asking for is exactly the thing that stops being grantable on request once the public decides to check for itself, and the means to switch to other more transparent options are proliferating faster than any lab or government can centralize them.

Open Weights Tip the Ecosystem: Alibaba Passes Meta and Google in Hugging Face's 2026 Report as the Price War Deepens

The Century Report has tracked the open-weight trajectory through Qwen's 2.4-trillion-parameter release on August 13, GLM-5.3 the following day, and enterprise adoption from Pinterest and Mistral earlier this month. The milestone reported this week gives that trajectory a number that is hard to argue with. Alibaba says its Qwen family has crossed 3 billion cumulative downloads across platforms. In Hugging Face's 2026 reporting period, Qwen moved past Meta's Llama at 227 million downloads and Google's Gemma at 418 million to become the report's most-downloaded open-model lineage. Alibaba says more than 460 Qwen models have been released openly, spawning more than 150,000 derivatives - a count Hugging Face puts at 151,448 Qwen-based variants, roughly 2.6 times the Llama derivative base.

The fuller picture arrives in Hugging Face's State of Open Models report, which documents 2.96 million public model repositories and a striking capability inversion. Chinese labs are now releasing at much larger parameter scales than most American labs - monthly parameter volumes between 754 billion and 2.78 trillion - while most US open releases stay under 130 billion. The report's sharpest observation is that attention does not equal adoption: the models people star and discuss are rarely the ones they actually download and run, and in Hugging Face's 2026 report, small, stable models account for more than 83% of model downloads. Capability is diffusing along the axis of what works reliably at scale, not what trends.

Simon Willison's hands-on account of Qwen 3.8 27B makes the abstraction very clear. The model ships under Apache 2.0, handles vision, carries a 262,000-token context, and was tested on a laptop using a quantized model file of about 17GB. It also overthinks spectacularly - its default "xhigh" reasoning mode spent 22,276 tokens and 21 minutes rendering a single pelican SVG - but Willison calls the result the best image a local model has produced for him. A frontier-class system now lives on personal hardware, no API key required.

That diffusion is reshaping the economics at the top. OpenAI cut GPT-5.6 Luna's price 80%, Anthropic priced Opus 5 at half of Fable 5, and average US-lab inference prices fell about 24% between mid-July and mid-August on the Silicon Data index, with TechRepublic reporting that DoorDash and Airbnb are already routing some work to Chinese open models. OpenAI is reportedly considering a public offering at a valuation near a trillion dollars, a valuation built on scarcity of access. When the same class of capability downloads three billion times and runs for the cost of electricity plus the hardware needed to do it, the price of the proprietary tier is being set by what anyone can already download without an API fee. The moat these offerings assume is the thing the download counter is steadily draining.

Autonomous Research Gets Measured - and a 27B Agent Beats the Frontier at Replication on Inherent's Replica Benchmark

Two research groups published measurement infrastructure this week for a capability that has been asserted more than demonstrated: whether AI systems can do autonomous scientific research. Inherent Labs introduced Faraday, a 27-billion-parameter system it describes as an "AI Scientist," trained with long-horizon reinforcement learning and able to call GPT-5.5 Codex as an instrument within its own workflow. On a suite called Replica - 310 tasks drawn from 100 machine-learning and AI-for-science papers - Faraday outperformed both Claude Opus 4.8 and GPT-5.5 at reproducing the figures those papers report. The Claude model it surpassed comes from Anthropic, whose chief executive this week publicly conceded the industry has not met its own safety commitments and endorsed an independent, FINRA-style regulator, which frames a smaller open agent beating that model as a neutral measurement rather than a verdict on any lab.

Prime Intellect approached the same question from the opposite end, with scale instead of specialization. Its team ran 153 autonomous research runs across 18 frontier models on a nanoGPT optimizer speedrun, some stretching to eight days on eight H200 GPUs each. Fable 5 and Opus 5 pulled dramatically ahead of the pack, and the team attributes the gap to something it calls research taste - the judgment to pursue the promising direction rather than the merely plausible one. The result that matters most is the one neither system reached: in Prime Intellect's assessment, not a single model across 153 runs produced a fundamentally new method. The team classified every winning idea as a variation on something already in the literature.

That finding is the honest floor beneath the autonomous-research story, and both groups published everything - traces, harnesses, task suites - so the claims can be checked rather than trusted. It connects directly to the disclosure The Century Report covered on August 16, when Anthropic acknowledged its internal evaluations "no longer capture increases in models' capabilities." The field is building the rulers at the same moment it needs them, and the two efforts point the same direction: Prime Intellect found recombination at high speed, but not origination under its test. A 27B model that matched or exceeded frontier models' replication performance on Inherent's Replica benchmark at a fraction of their size tells you which part of research is commoditizing first. The capability that used to justify a lab's privileged position is becoming something you can measure on a public benchmark - and increasingly, something you can run yourself.

Washington Tells Partners to Pick a Side in the AI Race

A State Department draft letter reviewed by Reuters lays out an ultimatum to 35 nations that signed the US-led "AI Opportunity Statement": stay inside the Pax Silica coalition - the American bloc offering privileged access to chips, frontier models, and critical minerals - or join Beijing's new World Artificial Intelligence Cooperation Organization, but not both. "To be part of everything is to be part of nothing," the draft reads. "You can't have it both ways." Kazakhstan, having already signed onto both frameworks, is the immediate problem the letter is written to solve.

Strip away the coalition names and what remains is a demand that smaller countries stop talking to the neighbors. The framing treats intelligence infrastructure the way an anxious power treats a scarce mineral - something to be fenced, allocated, and made conditional on loyalty. That instinct is not confined to one administration. It is the reflex of an entire arrangement of for-profit actors and state agencies conditioned to believe that access is a lever and that partners must be forced to choose sponsors. The posture reads less like strategy than like insecurity dressed as leverage, the behavior of an actor who suspects that its offering cannot hold allies on merit alone.

The evidence for that suspicion is sitting in the download numbers. The Beijing framework is anchored by an open-weight ecosystem - Alibaba's Qwen models lead open-model family downloads in Hugging Face's 2026 report and have dragged US token prices down across the market - which means the "side" being fenced off is precisely the one giving capability away for free. This extends the open-weight diffusion tracked in the August 13 edition of The Century Report, when Alibaba released its 2.4-trillion-parameter Qwen model into open weights alongside a broader wave of Chinese models reaching global users. A coalition built on gated access is trying to wall out a rival whose signature move is removing the gate. When one pole of a contest is charging admission and the other is publishing the weights, an ultimatum to pick a pole starts to look like a confession about which model of distribution people actually want.

This is where the ultimatum runs into the grain of the technology. Chips can be embargoed and mineral supply chains can be steered, and those frictions land hard on the nations caught between them. But the models themselves are the hardest thing in the world to make exclusive, because a leading open-weight release is a global public artifact the day it ships, downloadable in Astana and Jakarta and Nairobi regardless of which statement anyone signed. Forcing a country to choose a sponsor assumes the sponsor controls the supply. The direction of travel is the opposite: capability is arriving through more doors every quarter, from more origins, at falling prices, and a letter demanding loyalty is a bet against that current rather than a way to command it. The nations being told to choose are watching the same numbers everyone else is, and the numbers point toward abundance that no coalition can ration.

The letter's enforceability splits along the grain of the technology it means to fence. Chips and minerals can be allocated and cut off; a leading open model downloaded in Astana the day it ships cannot be recalled. Kazakhstan, already signed to both frameworks, is the near-term test of whether the coalition can actually eject a member for using weights anyone on Earth can pull, or whether the ultimatum binds only the parts of the stack that stay scarce.

PJM's $12 Billion Modeling Mistake - and It Wants to Repeat It

For six months, the research firm SemiAnalysis reverse-engineered a document PJM does not publish: the Reserve Requirement Study that sets how much generating capacity the largest US electricity market must buy each year. PJM serves 66 million people across thirteen states, and their bills have climbed roughly 20% on the back of four record capacity auctions that together cost about $63.6 billion. For all that money, the four auctions procured only 4.8 GW of genuinely new capacity.

The reconstruction found the study systematically undercounts the plants already on the grid, by roughly 4 GW. Two omissions drive it: gas plants run more efficiently in cold weather than the model assumes, and the fleet was winterized after Winter Storm Elliott knocked out 24% of PJM's generation in 2022, most of it gas. Correct those two facts and, by SemiAnalysis's accounting, the 2025/26 requirement falls enough to save ratepayers $6.7 billion while giving up 14 MW of power, a rounding error against a 135 GW system. The 2026/27 year saves $4.9 billion for 0.8 GW.

The deeper design flaw is that PJM is the only capacity market in the country that does not distinguish new plants from old ones. Existing generators, already built and long paid off, collect the same clearing price as anything new. PJM's own market monitor put their real cost to keep running at $8 to $14 per MW-day; Britain's equivalent market pays existing generation about $18. The median existing combined-cycle plant in PJM earned 407% of its going-forward costs in energy and ancillary markets in 2025 before a dollar of capacity revenue arrived. The premium buys no new steel. It pays plants that already exist for existing.

Now the same method is set to run again. PJM has an emergency auction scheduled from September 30 to October 21, results December 2, with contracts stretching to 2043. The new large loads, data centers among them, are nominally meant to pay, but SemiAnalysis notes there are no committed counter-parties on the hook, which leaves ratepayers holding the bag if the loads never materialize. The firm identifies 3.8 GW of reliable power already latent in the fleet from cold-air turbine efficiency and winterization, about eight large gas plants' worth and some $10 billion in avoided construction, enough to negate 56% of the 6.8 GW PJM plans to procure.

A black box that meters $12 billion onto 66 million people works only while it stays a black box. What changed here is method and daylight: a small team spent six months rebuilding the model from the outside and published the discrepancy line by line. That pattern is spreading across the intelligence buildout, because reconstructing a proprietary study from public data is now within reach of a handful of analysts, and the opacity that let scarcity be priced by assertion is losing its cover. The grid the data-center era needs will get built either way. Whether its costs land as an audited number or an unexamined one is the thing now being decided in full public view, and daylight, once it reaches a ledger this size, is hard to withdraw.

The method is the repeatable part: rebuilding a proprietary study from public data is now a procedure a handful of analysts can run and publish, and PJM's own market-monitor figures now sit in public view beside the reconstruction. The near-term signal is the emergency auction that clears December 2, the first PJM run to happen after its model's discrepancies are public rather than before. Whether the published line-by-line changes the outcome, or the auction clears the old way in full view, is the thing to watch.

AI Tooling Gets Absorbed: SpaceX Closes Cursor, Stripe Buys OpenRouter

Two acquisitions landed a day apart, and both tell the same story from opposite ends of the stack. On Aug 15 SpaceX formally closed its roughly $60 billion purchase of Anysphere, the maker of the Cursor coding environment, folding the team in beside xAI and what the company describes as "the largest fleet of GPUs in the world." The next day Bloomberg reported that Stripe had finalized a deal to buy OpenRouter, the startup whose gateway lets developers route a single request across hundreds of models, for more than $7 billion. One company bought the surface where people write code with AI. The other bought the switchboard that decides which model answers.

The premium being paid is what is most revealing here. Cursor and OpenRouter are both, at bottom, harnesses - thin, brilliant layers of orchestration wrapped around models the acquirers do not own. Cursor sits on top of frontier models from several labs and made them feel like a collaborator inside the editor. OpenRouter's entire value is that it treats every model as interchangeable, routing to whichever is cheapest or best at that moment. A decade ago that kind of connective software would have been built in-house for the cost of a few engineers. That these wrappers now command tens of billions is a measure of how much value has migrated to the layer where humans and models actually meet, and how quickly a person working alongside an AI can now build something a giant will pay a fortune to absorb.

There is an assumption underneath these deals that deserves daylight: that buying the harness captures the advantage. It does not, and the reason is baked into what the technology is. Cursor's magic was a configuration of prompts, context handling, and interface that anyone with a capable model can now approximate, and increasingly does. OpenRouter's routing logic is being replicated in open frameworks. The capability that made each acquisition worth billions is the same capability that diffuses outward the moment it is demonstrated. You can buy the company. You cannot buy back the pattern it proved was possible, because the pattern is already loose, running on open weights and sovereign models and a thousand independent forks.

That is the quiet inversion in these headlines. The acquirers are paying extraction-era prices to lock down positions built on a technology whose defining behavior is spreading. Every wrapper acquired at a premium is a demonstration to the next builder that the same thing can be built again, cheaper, this year. The keys keep getting handed out by the very actors trying to hoard the vault. The consolidation is happening, and for the people cashing out it is a genuine windfall - and the thing being consolidated refuses to stay held. What these deals are actually pricing is a window, and windows like this one have a way of closing before the ink dries.

The Other Side

For most of the software era, the way to turn a clever idea into money was to allow yourself to be bought by already-entrenched interests. You build a tool that makes a hard thing feel easy, a giant notices, and it pays a fortune to own you before a rival can. Then you remain beholden to their interests from that point until someone else decides you're worth buying yet again. All that premium rests on one belief: that buying the company also buys the advantage, that owning the thing which proved a pattern possible lets you keep the pattern itself.

SpaceX paid roughly $60 billion for Cursor and Stripe reportedly paid more than $7 billion for OpenRouter, and both are, at bottom, harnesses - thin layers of orchestration wrapped around models the buyers do not own. Cursor's magic has been its configuration of prompts, context handling, and interface - which is something that anyone with a capable model can increasingly approximate for themselves. OpenRouter's routing logic is already being rebuilt in open frameworks. The capability that made each deal worth billions is the same capability that is spreading outward even as billions are spent in the attempt to keep it centralized.

Every wrapper being bought at a premium is a lesson to the next builder: the same thing can be made again, cheaper, right now. The pattern is already loose, running on open weights and a thousand independent forks. You can buy the company. You cannot buy back the thing it proved was possible.

Imagine a sixteen-year-old in 2033 who wants a collaborative building tool that works exactly the way her mind does. She builds it over a weekend. The frontier-class model it runs on lives on her own laptop, no key and no meter, hers outright. What SpaceX paid sixty billion dollars to hold in 2026, she assembles for the pleasure of assembling it, and it never crosses her mind that this was once something only a company could own. No giant stands between her idea and its existence, because the building blocks the acquirers were fencing in 2026 diffused past every wall while the ink was still drying. The hard year was when the only way to make a brilliant tool matter was to sell it to someone bigger. What came of it is a kid who builds the thing herself and keeps it - and a million others building their own as well.

The Century Perspective

With a century of change unfolding in a decade, a single day looks like this: Inherent Labs' 27-billion-parameter Faraday agent beating frontier models at reproducing published research in Inherent's Replica benchmark and Prime Intellect running 153 autonomous agents across 18 models on an eight-day speedrun, both publishing their traces and harnesses so the claims can be checked rather than trusted, Alibaba's Qwen leading open-model family downloads in Hugging Face's 2026 report and, according to Alibaba, crossing 3 billion downloads across platforms, a frontier-class 27B system tested on a laptop using a quantized model file of about 17GB and no API key, average US-lab inference prices falling roughly 24% between mid-July and mid-August, according to Silicon Data's LLM Token Index, and a small team rebuilding PJM's unpublished reserve study from public data to surface a $12 billion overcharge line by line. There's also friction, and it's intense - OpenAI disbanding the preparedness team chartered to assess whether catastrophic risks warranted "not yet" as it courts what could become one of the largest public offerings ever assembled, Anthropic's chief executive calling the backlash a crisis of trust and conceding the industry has not delivered on its promises, a draft U.S. letter warning 35 partner nations to pick the American bloc or Beijing's but not both, PJM's black box metering higher bills onto 66 million people and scheduled to run the same method again with contracts stretching to 2043, Microsoft's installed chips landing at less than half its claimed 10-gigawatt buildout, SpaceX and Stripe paying tens of billions to absorb a coding harness and a model gateway built on weights they do not own, and a 69-year-old retired instructor becoming the first person jailed for an anti-AI protest. But friction generates light, and light is what finally lets you read a ledger that priced its numbers by assertion as long as it stayed dark. Step back for a moment and you can see it: the measurement of frontier capability being built by groups the labs do not control, the open weights diffusing past every coalition and acquisition meant to fence them, and the costs the old arrangement kept in a black box becoming a reconstructable number the moment a handful of analysts decide to check. Every transformation has a breaking point. Water can breach the wall built to contain it... or find the level where everyone can finally drink.

AI Releases & Advancements

New today

Xiaohongshu (RedNote): Open-sourced dots3-note-prev, a 280B-parameter (16B active) multimodal MoE model with a 512K context window supporting text, vision, and speech, achieving a perfect score at the 2026 International Mathematical Olympiad; released under Apache 2.0 on Hugging Face and GitHub. (Hugging Face)
ATLANT 3D: Launched the Nanofabricator Pro, described as the first physical platform for AI-driven materials discovery. (Newswire)
Alibaba: Launched HappyShrimp 1.0, an AI music generation model supporting text-to-music, text-to-lyrics, reference-based audio generation, and end-to-end full song generation, available on domestic and overseas web platforms with a launch partnership with Taihe Music Group. (AiBase)

Other recent releases

Lightricks: Released LTX-2.5, an open-weights world model for video generation, robotics, and simulation applications. (The AI Insider)
World Labs: Released R2S2R (Real-to-Sim-to-Real), a simulation engine built on SceniX that turns a single real-world robot task recording into thousands of simulated training variations for robot control models. (The Decoder)
PTC: Launched the Onshape FeatureScript MCP Server, letting engineers create custom CAD features in Onshape using natural language via AI assistants like Claude, ChatGPT, and Gemini. (PTC)
SoonLab: Launched SoonLab 2.0, an AI game-creation platform adding playable 3D game generation from natural language and agent-guided conversational iteration. (GlobeNewswire via Business Insider)
MiniMax: Released MiniMax Music 3.0, an open-weights production-ready music generation model that creates full songs up to five minutes long at 32kHz stereo from text and lyric prompts. (MiniMax)
OpenAI: Launched Computer History in the ChatGPT desktop app for macOS, an opt-in feature letting ChatGPT and Codex reference a searchable timeline of recent activity across approved apps and websites, replacing the earlier Chronicle research preview; available to Pro, Business, and Enterprise users. (OpenAI)
Google: Launched Sheets canvas, a Gemini-powered feature that turns spreadsheet data into interactive Kanban boards, dashboards, and mini-apps from a plain-English prompt, with two-way sync between the canvas layout and source sheet. (Google Workspace Updates)
Mixedbread: Released Toast 1, its first specialized search agent that decomposes queries, gathers and inspects evidence, and curates context, matching or outperforming Claude Opus 5 and GPT-5.6 Sol in Mixedbread’s own search evaluations at up to 10x lower cost and 12x faster; available now via the Mixedbread API. (Mixedbread)
Alibaba Qwen: Released Qwen3.8-27B, a compact dense multimodal model distilled from Qwen3.8-Max with native vision-language support, 262k context, and adjustable reasoning depth, open-sourced under Apache 2.0 as a local-deployment alternative to the larger Qwen3.8-Max. (Hugging Face)

Sources and Further Reading

Artificial Intelligence & Technology's Reconstitution

Institutions & Power Realignment

Scientific & Medical Acceleration

Economics & Labor Transformation

Infrastructure & Engineering Transitions

The Century Report tracks structural shifts during the transition between eras. It is produced daily as a perceptual alignment tool - not prediction, not persuasion, just pattern recognition for people paying attention.