In February 1950 a Weather Bureau statistician named Glenn W. Brier sent the Monthly Weather Review a three-page paper with a dry title, "Verification of Forecasts Expressed in Terms of Probability." Brier had noticed that forecasters graded by a scoreboard were tempted to "hedge" or "play the system," because "it is often easier to analyze the effect of different possible forecasts on the verification score than it is to analyze the weather situation." His fix was to ask for probabilities, wait for the weather, square the gap between each probability and what happened, and average. Brian Roemmele uses that small formula to judge a new kind of AI: Drex, a family of open-weight "decision models" that Nace AI described in a technical report on November 6, 2026.

How the score works. A forecast of 80 percent rain scores 0.04 if it rains and 0.64 if it stays dry, so confident mistakes cost the most. Lower is better. The common yes-only version runs from 0 to 1; Brier's original adds the squared errors for every outcome and runs from 0 to 2, exactly double in a yes/no problem. The key property is that honesty pays. If the true chance is 80 percent, saying 80 has an expected score of 0.16, while saying 60 or 100 both score 0.20. Over time you cannot beat the rule by stating anything other than what you really believe. Roemmele warns that the math cannot stop a dishonest record keeper who cherry-picks questions, hides forecasts or moves deadlines.

Brier's own ten forecasts. In Brier's Table 1, rain came on three of ten occasions. Two misses dominate: a 70 percent call on a dry day and a 40 percent call on a wet one. The yes-only score is 0.095 and the original score is 0.19, which matches the paper. Always forecasting 30 percent would have scored 0.42, so the forecaster was clearly adding skill. Roemmele separates three ideas: accuracy (how often the top pick wins), calibration (whether 80 percent calls come true about 80 percent of the time across many cases) and the Brier score, which rewards both.

What Drex is. Each Drex model takes a context, a question and a fixed list of allowed options, and returns one probability per option. It cannot invent an answer that is not on the list, though it can still pick the wrong one. The family is Drex DLM (8 billion parameters), v1.5 (9B), v2.0 (35B total with 3B active, a mixture-of-experts design) and Drex Max (27B). Max is built on Qwen3.8-27B, can read images and uses a special readout head, so it will not score correctly in an ordinary chat pipeline. Roemmele advises adding a "review required" option and a separate policy layer, because choosing a button is not the same as having permission to press it.

The numbers, all reported by Nace and not re-run by Roemmele. Max scores 67.52 on the public Decision Index 0.4.1, against 60.70 for Jev 1.14.0, and 66.60 with private tests added. Nace cites an Artificial Analysis score of about 64 (63.7). On Decision Bench v4 with raw images it gets 94.0 percent (1,007 of 1,071), statistically tied with d3-turbo and GPT-6 Sol and 1.6 points behind Claude Opus 5.5. The headline speed of 183 milliseconds is a separate launch measurement; on Decision Bench v4 the median is 0.66 seconds, against 2.05 for GPT-6 Sol and 4.33 for Opus. Using the actual image instead of a text description raises image-question accuracy from 63.3 to 98.3 percent (38 versus 59 of 60). Shuffling the option order changes the top answer 4.3 percent of the time, and reversing it does so 6.2 percent of the time.

The 93 percent claim. Nace says that when Max is at least 93 percent sure, it is right 97 percent of the time. Roemmele calls this a starting point. Overall, Max is mildly overconfident: 87.9 percent accuracy against an average top probability of 91.2 percent, and the smaller Drex v2.0 is better calibrated. Keeping the most confident 80 percent of answers lifts accuracy to 96.1 percent, and keeping the top half lifts it to 99.3 percent. He also flags a trap: the API's "confidence" field is a chance-adjusted transform, not the raw probability.

The Rain Log. Roemmele publishes a small standard-library Python script that reads a CSV of id, p_yes and outcome, prints both Brier conventions and a ten-bucket reliability table. He ran it on Brier's data and got 0.095 and 0.19. A synthetic, perfectly calibrated 200-row set still scores 0.1675, because calibration does not make uncertain events certain. He also offers a "Brier's Desk" prompt that makes any chat model log scoreable forecasts and write "Calibration unverified" when it has no history. His deployment advice: use separate data for building, threshold-picking and testing; track coverage and missing outcomes; and require human approval for irreversible actions whatever the confidence.

Licences, hardware and caveats. Max, v2.0 and v1.5 use Open RAIL-M terms, DLM is non-commercial CC BY-NC 4.0 and the code is Apache-2.0. A Q4 text-only build of Max runs on an Apple silicon Mac with 24 GB or more, and Q8_0 needs 48 GB. Roemmele did not run Drex itself. His business ideas (calibration audits, ticket and claims triage, local models for privacy-bound firms, scored-forecaster newsletters and decision-log tooling) are explicitly not financial advice. He closes on Brier: a rain gauge does not change the weather, but it gives an honest reading, and a decision log can do the same for an AI system.

In 1988 Pentti Kanerva published Sparse Distributed Memory from NASA Ames Research Center under contract NCC2-408. MIT Press issued it as a Bradford Book.

The design treats addresses as long binary strings. In the account that became standard, those addresses run 1,000 bits, a space of 2^1000 patterns, while only about a million physical locations are ever built.

A write places the same pattern into every hard location that falls inside a Hamming-radius neighborhood of the given address. A read presents a cue, activates the same neighborhood, and lets the stored traces vote by majority. A cue that is only close is enough. In the early demonstrations a noisy circle cleaned itself toward a stored prototype. A sequence stored as successive Roman numerals continued forward when cued from the middle.

The same year a Stanford-NASA prototype used 256-bit addresses and from 8,000 to 128,000 ordinary DRAM locations. The physical size of the memory was fixed. What it could recall was not.

On November 5th 2026 Tushaar Gangavarapu, Gil Tabak, and Clemens JS Schaefer posted Muir on arXiv. The name points to the redwoods at Muir Woods whose rings keep centuries of history in a fixed circumference. Their 1.34-billion-parameter model reads streams of ten million tokens from a working state of 25 megabytes in float32. The authors calculate that a Transformer of the same size would require a 2.1-terabyte bfloat16 key-value cache to attempt the same stream; their own Transformer baseline ran out of chip memory at 128,000 tokens.

The model was trained on sequences of 4,096 tokens. On author-reported tests its next-token score held without drift out to the ten-million-token window. A static-readout variant recovered a hidden passkey near 100 percent of the time at one million tokens (about 90 percent with no extra training, 99.4 percent after a short 64,000-token cooldown).

The earlier Gated DeltaNet-2 design collapsed past 4,000–8,000 tokens. Training throughput matched the gated-delta baseline and ran about 2.7 times faster than the Transformer at 32,000 tokens. All of these figures are the authors’ measurements. The weights have not been released. The work has not yet been peer-reviewed. The largest model remains 1.34 billion parameters.

Muir belongs to the linear-attention family that folds every token seen so far into a fixed matrix. Slow-decaying rows of that matrix accumulate large norms and drown quieter traces from earlier in the stream. The usual remedy is stronger erasure at write time, which discards the quiet traces. Muir’s matched-filter readout multiplies the query by a learned gain that re-weights the rows before the inner product. Channels that have been growing loud are dialed down; channels that still carry distant signal are dialed up. The first author described the gate the next morning on X as the fix for “100x slow-channel glare” and noted that the model “Streams 10M tokens in a 25MB state.” The paper appeared on the main AI research feed the same day. Recaps were still being updated on 8 November.

I have not seen another modern language model treat Kanerva’s fixed-size specification as an explicit engineering target. The correspondence is not identity. Kanerva worked in binary vectors and Hamming distance; Muir works in continuous channels and learned gates. The shared constraint is the one that matters for machines you own: the memory does not grow when the document grows.

The Fixed Desk is a short local test that measures whether the model on your machine obeys that constraint. Muir’s weights are not public, so the test uses open cousins that already keep a recurrent state, among them the Qwen3 hybrid variants that mix gated-delta layers with ordinary attention. Load a long document in successive windows.

Log the process resident set and the size of any explicit key-value cache after each window. A model that follows the Fixed-State Rule holds the resident set within a few megabytes of its starting value once the window length exceeds the training context. A model that fails the rule shows a linear climb.

The same run also runs the Passkey Check: hide a short unique string near the beginning of a multi-million-token file, then ask for it after the rest of the file has been read. Recovery rate is the only score that counts.

Kanerva’s Desk SuperPrompt turns any long-context model into a write-erase-retrieve journal. At each step the prompt carries three fields: the current compact state, the new observation, and the instruction to erase the prediction that the current state would have made for this observation before writing the new value. Retrieval is performed by similarity to the stored state rather than by position. The prompt is short enough to sit inside the context of models that already run on a laptop.

When a library fits in a fixed 25 megabytes the economics of local archive agents change. Document question-answering no longer prices itself on growing cache. A year of mail, a code repository, or a scanned microfilm run can be held in the same RAM footprint that already stores the model weights. The opportunity is ordinary: agents that answer from the whole archive without renting the next larger cloud instance every time the archive lengthens.

The limit is the one already stated. The published numbers are author-reported, the weights are unreleased, and the largest published model is still research scale.

The Fixed Desk measurement on the machines you already own is the step that does not wait for those weights.

Heather Cox Richardson opens her November 6 letter with a blunt reading of Tuesday's midterms: Americans rejected the Trump administration. In her telling, voters turned away from a strongman who ignored the Constitution, undermined democracy and enriched himself and his allies, and they did it by a margin of more than 12 million votes — demanding an end to corruption, to the war on Iran, to costly tariffs and to masked immigration agents seizing people off the streets.

She argues the numbers speak for themselves. Democrats won the national vote by more than 13 points, which she frames as a vote for the rule of law, for the Framers' checks and balances, and for policies that make food, housing and health care affordable rather than steering money to billionaires. The Economist, she notes, called it "the biggest congressional landslide in 50 years." By contrast, Trump's 2024 win over Kamala Harris — about 2.3 million votes and under 50% of the total — now looks like a fizzle next to the "mandate" he claimed.

Richardson lingers on who won. Voters elected Black Democrats Keisha Lance Bottoms in Georgia, David Crowley in Wisconsin, Aaron Ford in Nevada and Stacey Plaskett in the U.S. Virgin Islands, along with Black Republican Byron Donalds in Florida. Michigan's Abdul El-Sayed becomes the nation's first Muslim senator, and Rep. Ilhan Omar — a frequent Trump target — was reelected with 76.5% of her district's vote.

The coalition shifted too. Citing the Associated Press, she reports that Hispanic voters swung hard back toward Democrats, with seven in ten Hispanic Texans backing Senate candidate James Talarico. For the first time since 2008, Democrats won every House district touching the U.S.-Mexico border, and fourteen new Hispanic Democrats are headed to Congress. Black voters, especially in the South, turned out in unusually high numbers; the New York Times found about 80% of Black votes in House races went to Democrats. Since the election, she says, panicked Republicans have been talking about regrouping, ending the Iran war and focusing on affordability.

Trump, however, went the other direction. After appearing dazed by the result, he fired chief of staff Susie Wiles on Friday and replaced her with Stephen Miller. Richardson describes Miller as a white nationalist who appears to embrace "great replacement" thinking, the driver of the first-term Muslim ban and the second-term mass-deportation push, including the 3,000-arrests-a-day target she links to the deaths of U.S. citizens Renee Goode and Alex Pretti. She says Miller championed rule by emergency declaration, urged invoking the Insurrection Act, and — per NBC's Mike Hixenbaugh — pushed a plan to institutionalize mentally ill homeless people that led the Justice Department to stop enforcing the 1999 Olmstead disability-rights decision and drop at least five related cases, a signal to states facing Medicaid cuts. Citing Phillips O'Brien, she adds that Miller sides with Russia on Ukraine, and that his promotion suggests Trump wants to deliver what he can for Putin before the new Congress begins investigating.

She lists the circle Trump is tightening around himself: Defense Secretary Pete Hegseth, aide Natalie Harp and budget director Russell Vought, a Project 2025 author. Democrats reacted sharply — Rep. LaMonica McIver called Miller "a white supremacist," Chuck Schumer said Trump "learned nothing from his historic defeat," and Hakeem Jeffries said Miller "has American blood on his hands." One House Republican, told of the pick by Politico, asked if the reporter was joking.

Richardson reads the move as a warning to Republicans as much as Democrats: Axios reports Trump plans to keep fundraising to control 2028 candidates, and the Wall Street Journal reports the White House told more than 3,000 conservative groups it is "monitoring" their statements. "It all feels very desperate," she writes. She closes with Trump at a NASA event Friday night, in a campaign hat, rambling about nuclear spacecraft to Mars, a return to the Moon and his MIT uncle, then telling a silent room that he wasn't on the ballot — "because if we were on the ballot we would've won."

In his victory speech on November 3, after winning a U.S. Senate seat in previously Republican Texas, Senator-elect James Talarico framed the result as something older than party: Americans joining to keep wealth and power from being taken for a few. Elon Musk had spent at least $225 million in the midterms trying to keep Texas government in Republican hands. Voters elected a Democratic governor, lieutenant governor, attorney general, and state house anyway.

“We just took back Texas,” Talarico told the crowd. Money was no match for the people of Texas, he said; a politics of division was no match for a politics of love. The message to billionaire and trillionaire megadonors, and to everyone defending a broken system, was that power belongs to the people.

He leaned on his faith, his years as a middle-school teacher of underprivileged students, and his childhood as the son of a single mother. Scripture and his own life, he said, taught him to “always bet on the underdog.” The campaign had gone up against nearly 40 years of doubt and more than $300 million — more than any Senate campaign in American history — plus negative ads, smear campaigns, foreign bots, and AI deepfakes. Those attacks, he said, were “the final howls of a dying establishment.”

Talarico said the movement united first-time voters, three-time Trump voters, and working people from every corner of the state. The win was for teachers and nurses, farmers and ranchers, young people trying to buy a home, and parents trying to put food on the table. “This win is for every Texan who’s ever wondered if anyone is listening. I heard you.”

He told the crowd the result was bigger than either party. America’s economic system, political system, and even relationships with one another felt broken, “because the most powerful people in the world want it that way.” He rooted his politics in the commandments to love God and love neighbor, and said people still want the same things: a safe neighborhood, a good job, a well-funded public school, and a doctor when they need one.

The billionaires who own social-media algorithms, cable news networks, and the politicians fighting on screens, he said, want Americans at one another’s throats — looking left and right instead of up. “The real fight in this country is not left versus right. It is top versus bottom.” Division by party, race, gender, religion, and culture, he argued, hides pocket-picking: defunded schools, gutted healthcare, and higher taxes on everyone else while the people at the top cut taxes for themselves. “It is the oldest strategy in the world, divide and conquer. But tonight we sent a crystal clear message to the people at the top. We will not be conquered.”

He said the country is tired of ten years of politics as blood sport, trolling, and professional wrestling — the kind that tears families apart and ends friendships. People across the spectrum, he said, want a politics of love rather than hate, fear, or division.

He described how Democrats won Texas: 60,000 volunteers in every corner of the state delivered bottled water to Sand Branch, a community south of Dallas with no running water; donated school supplies to an elementary school in El Cenizo, a small town on the border; and collected more than 20,000 pounds of groceries for Fort Worth families squeezed by the cost of living. The world changes person to person and neighbor to neighbor, he said, before it shows up on television.

The work ahead, in his telling: get big money out of politics so no billionaire or trillionaire can buy elections; unrig an economy that only benefits the top 1 percent; pass universal health care; invest in teachers and students; end “corrupt forever wars” that drive up prices and kill service members; and heal what he called the spiritual corruption at the root of those crises. “That old politics is dying, and a new politics is being born through each of us.”

Talarico won by 12 points. Richardson writes that he correctly identified how local, relational organizing built the communities that elected Democrats in 2026, and that this kind of politics threatens establishment politics, especially MAGA. President Trump, she notes, reinforced his need for division the same day by offering the White House press secretary job to Dana White, the UFC’s chief executive, whom he described as a “violent Fighter and Warrior,” and by encouraging White to bring along UFC fighters. “We need these people to defend us against the Lunatic Left,” Trump posted.