Sep 14 close · +40% vs pre-crisis ~$76

Sep 15 · new all-time record ($6.2694) — sixth straight · +68% vs pre-war $3.72

Sep 15 · +18¢ in a week (AAA) · +54% vs Jan $2.81

Sep 4 · down 1.2M in a week · down 130.1M from pre-war 415.4M · lowest since Dec 1982

Sep 4 · up 2.1M in a week · 13% below 5-year average · East Coast stocks 28% below last year

These charts show how oil, gasoline, diesel, and natural gas prices have changed. Brent is a widely used benchmark for the price of crude oil. The gasoline and diesel charts use US national averages; the natural gas charts cover Europe and Asia. Every plotted value comes from a source. Missing readings are left out.

The national average has risen from $2.81 in January to $4.33, an increase of 54%. Prices peaked at $4.50 in May. Every month since March has averaged above the levels seen before the crisis.

Diesel cost $3.72 in the last week before the war, on Feb 27. By Sep 15, it had reached $6.27, an increase of 68% and its sixth consecutive record. It passed the previous record of about $5.85, set in June 2022, on Sep 4. This chart shows selected readings. You can see the daily Sep 3–15 readings in the price spread chart below. Diesel prices have risen faster than gasoline prices as supplies of refined fuel have tightened.

TTF, $/MMBtu · filled: EIA weekly futures avgs (Jan–Apr) · hollow: individual contract quotes (May–Sep) · different instruments — a trend, not one series

TTF is the benchmark used to track natural gas prices in Europe. It traded above $28 on Sep 10, eased to $27.00 on Sep 11, and stood at about $27.80 on Sep 14 — the highest levels since December 2022. That's roughly 153% above the price before the strait closed. The Sep 14 value is converted from €81.98 per megawatt-hour to the dollar units used in this chart. Prices have risen as unplanned maintenance at Norway's Asgard and Troll fields and tanker attacks put pressure on supplies.

By January 20, European gas storage had fallen to 48%, compared with a five-year average of 63%. Running that low left Europe buying liquefied natural gas (LNG) for immediate delivery during the season the strait closed.

JKM, $/MMBtu · filled: EIA weekly futures avgs (Jan–Apr) · hollow: assessed spot (May–Sep) · different instruments — a trend, not one series

JKM tracks the price of liquefied natural gas (LNG) delivered to Asia. It reached the high-$28s on Sep 10 — its highest level in roughly two and a half years — and held around $28.50 on Sep 11, about 167% above the price before the closure (JOGMEC). Damaged production units at Qatar's Ras Laffan complex — about 17% of the country's LNG export capacity — are expected to be offline for 3–5 years, forcing Asian buyers to look elsewhere for supplies. With little gas in storage, the region is particularly sensitive to changes in supply and weather.

Brent: sourced points (EIA monthly avgs: Mar $103.0 · Apr $117.29 · Jul $83.76) · WTI: weekly (FRED)

Brent rose from $76 before the closure to an intraday peak of $126 in March, then fell to $95 by mid-April and $74 on Jun 30 as hopes for a ceasefire grew. It passed $100 again on Sep 9 as tanker attacks escalated, rose $6.82 to $108.03 on Sep 10, and settled at $104.61 on Sep 11.

Brent closed at $105.68 on Sep 14 after reaching $106.73 earlier in the day. Prices rose after talks between Gulf foreign ministers and Iran were postponed. Reuters also reported that stocks at Yanbu, Saudi Arabia's Red Sea port, could support another five to seven days of exports if the East–West pipeline remains closed.

The gray line shows West Texas Intermediate (WTI), a US crude oil benchmark. Most WTI readings come from FRED's weekly spot series; the Sep 10, 11, and 14 readings are futures closes. WTI is also the crude price used in the fuel price spread chart beside this one. Hollow markers indicate a source's rounded estimate.

Retail price less WTI, $/bbl · weekly (EIA/AAA − FRED).

This chart subtracts the price of crude oil from the retail price of fuel, with both expressed in dollars per barrel. The difference covers refining, transportation, and retailing. It helps explain why prices at the pump can keep rising even when crude gets cheaper. Diesel's spread grew from about $89 in January to $161.9 on Sep 15. It briefly narrowed to $148 on Sep 10, when crude prices rose faster than pump prices. Gasoline's spread has also widened, from about $59 to $78–84. This uses retail prices from EIA and AAA, minus WTI from FRED, with futures closes for Sep 10, 11, and 14. The EIA's official crack spread uses wholesale fuel prices, so the values differ. Eight fuel readings have no same-day WTI value; those use the nearest trading day's price, within 1–2 days. No missing prices are estimated.

When the world uses more oil than it produces, the difference comes out of storage. These charts show how much oil has been withdrawn, how much is left, and how long the US emergency reserve could last under each of the model's three scenarios.

Before the war, the world produced about 4 million more barrels of oil per day than it used. Since the war began, it has had to draw on stored oil every month. The IEA's latest report estimates a full-year supply loss of 5.7 million barrels per day, about 6% of the world's oil, and expects Middle East oil flows to remain below normal until 2027. Global production fell to 100.1 million barrels per day in August, with more than 10 million barrels per day of Gulf production still shut down. Saudi production alone fell by 2.3 million barrels per day that month, to 5.97 million.

World oil balance, million b/d — production minus consumption, EIA STEO Table 3a, reported months (Jan–Aug actuals; the EIA's forecast tail is not shown) · the physical loss peaked at 11.2M b/d of Gulf shut-in in May — demand destruction and non-Gulf supply absorbed most of it · the IEA's observed-inventories count: 507 mb drawn since February — 2.8 mb/d on average, 95 mb of it in August alone.

Withdrawals and changes in demand, using figures available as of Sep 11.

year-to-date · EIA est. (Sep 9)

US Strategic Petroleum Reserve

pulled from 32 countries · IEA

withdrawals inferred from customs data · official SPR untouched

full-year 2026, cut from −1.6 in the August edition · IEA OMR, Sep 11

Q3 2026 forecast — supply below demand · IEA OMR, Aug 12

EIA weekly ending stocks, million bbl

The Strategic Petroleum Reserve (SPR) is the US government's emergency supply of crude oil. It was created after the energy shortages of the 1970s. The reserve held 415.4 million barrels when the war began. The latest report puts it at 285.4 million as of Sep 4. Oil is being released to help make up for supplies that can't leave the Gulf. You can use the withdrawal rate to see how quickly the emergency supply is being used.

The red lines mark the model's reserve thresholds, or floors. At about 300M barrels, some caverns risk damage and can't safely be refilled after a withdrawal. The first report below that threshold was for the week ending Aug 7, at 298.7 million barrels.

The other floors are 250M, the GEF minimum for sustained withdrawals; 180M, the hard operating limit; and 70M, the Department of Energy's stated safe minimum. The model stops withdrawals at 70M.

The dashed lines show what happens if withdrawals continue at 0.45M, 0.70M, or 1.35M barrels per day from the Sep 4 level. The estimates extend to about March 2027.

Withdrawals slowed to about 0.2 million barrels a day in the week ending Sep 4, while diesel stocks rose by 2.1 million barrels, according to the EIA's Sep 10 report. The next report is due Sep 16 and covers the week ending Sep 11.

Estimated odds, updated when specified events occur.

The Saudi bypass pipeline was suspended, and the Houthis held the entire Red Sea coast. An official pipeline restart would return the odds to 10/50/40.

Tanker losses reached 10 per week, Brent passed $100, and Jazan was affected.

A shipping exclusion zone was imposed, and a base in a third country was hit for the first time.

Reported Hormuz traffic of 8.6M barrels per day was not backed by vessel tracking, which showed 7% of normal transits.

Initial model: about 65% odds of de-escalation.

Tankers can pass through Hormuz under an Iran–Oman agreement or with US escorts. Traffic gradually returns to normal over one to two quarters.

In this scenario, Brent moves toward $70–80 and reserve withdrawals slow to about 0.45M barrels per day. Stored oil lasts longer.

The war continues at its current intensity. Tanker attacks and shipping restrictions persist, some Iranian infrastructure remains offline, and the damaged Saudi bypass has no restart date. The strait remains partly open.

Brent stays in the $95–125 range, and reserve withdrawals run at about 0.70M barrels per day. Global stocks keep falling, with shortages developing later.

The disruption becomes a sustained closure or the fighting escalates. Tanker losses rise, shipping restrictions remain, and the bypass, Abqaiq, and Jazan stay offline for months.

Brent rises above $130, and reserve withdrawals reach 1.35M barrels per day. Shortages spread from the US East Coast to Russia, Europe, China, and aviation fuel.

These odds are based on judgment. They change when specified events occur, such as a pipeline restarting or a shipping agreement breaking down. A quiet week alone doesn't change them. You can read the rules and the events being watched on the model page.

This chart puts the current reserve in perspective. It held 727 million barrels at its December 2009 peak and 294 million at the previous low in December 1982. It held 285.4 million on Sep 4. The green line marks the level before the war. In the model, withdrawals stop at the 70 million barrel floor.

Large withdrawals began on Apr 3, when the reserve held 413.3 million barrels. They reached about 1.2 million barrels per day in May. In the week ending Sep 4, withdrawals averaged about 0.18 million barrels per day, down about 60% from the previous week. That's below the 0.45 million barrels per day assumed in the corridor-holds scenario.

US diesel and heating-oil stocks were 13% below their five-year average on Sep 4, according to the EIA.

−130.1M (−31%) since pre-war 415.4M

Lowest since Dec 1982 · down 1.2M barrels in the week ending Sep 4

Up 2.1M barrels in the week ending Sep 4; 13% below the 5-year average (EIA summary)

East Coast stocks are 28% below last year

At the 5-year average (EIA summary, week ending Sep 4)

Refined fuels remain in shorter supply than crude oil

The latest withdrawal rate and estimated dates for reaching the reserve thresholds, using the EIA's Sep 10 report for the week ending Sep 4.

4-week average, EIA weekly report (week ending Sep 4) · latest single week: 0.18.

Weekly withdrawals: about 9M barrels at the late-May peak, falling to 1.2M by Sep 4. The earlier 9.9M peak in the week ending May 15 falls outside this 16-week chart.

Estimated date at 250M barrels

The corridor-lapse scenario has the highest odds, at 50% as of Sep 11. It assumes withdrawals of 1.35M barrels per day from the reported 285.4M barrels on Sep 4. The standoff scenario, at 40%, reaches the same threshold on Oct 24.

Next floor — the 180M operable limit: ≈ Nov 21, 2026 on the lapse path, Feb 1, 2027 on the standoff path.

Crude oil has to be refined before it can be used as diesel, gasoline, or jet fuel. That makes refinery capacity just as important as the amount of oil available. US refineries are processing more oil than last year, while processing has fallen elsewhere and strikes continue to damage Russian refineries. The IEA describes the global refining system as “stretched to the limit.” Atlantic Basin refining margins reached records in August, led by diesel.

Utilization, % of operable capacity, weekly.

US refineries have operated above 95% of their available capacity every week since Jun 5, reaching 98.0% in the week ending Aug 28. The comparable period in 2025 averaged 90.8%. Fuel exports also reached a record 8 million barrels per day in August, according to OPEC. US plants are working close to capacity, but shortages in Europe, Asia, and Russia continue to put pressure on fuel supplies.

EIA Weekly Petroleum Status Report (WPULEUS3), week ending Friday · utilization = gross inputs ÷ latest reported operable capacity (EIA's definition) · 2025 line = same Jan–Sep window · in mb/d: runs 16.3–17.3 (STEO 4a), above 2025 in every month · see the fuel price spread chart in Prices for the effect on costs.

Refineries elsewhere are processing less oil. These figures from the IEA's August and September reports show the size of the decline.

summer peak, −4.2 mb/d below a year ago (OMR, Sep 11)

IEA forecast vs 2025 (OMR, Sep 11)

the quarter's further cut (OMR, Aug 12)

estimated capacity remaining, % of pre-strike

Strikes on Russian refineries are reducing the amount of fuel available to other countries. The chart shows about 75% of capacity remaining by mid-April and about 70% by Aug 29, according to the Moscow Times. The red bar shows how much current estimates differ: Ukraine's General Staff puts the capacity lost at 42.74%, Russian Forbes at 54%, and the IEA at more than 20%. You can compare these estimates with the reported outages and export restrictions below.

Aug 29, Moscow Times — up from ~25% in April

Early September estimates of capacity offline: 42.7% (Ukraine's General Staff) to 54% (Forbes)

record month, near-daily (Bloomberg, Aug 29)

~400K b/d, its only NW plant, two strikes in a month (UA.NEWS, Sep 2)

Ryazan (Rosneft) — Moscow's main supplier

~156K barrels/day; both primary units offline since Sep 6, with repairs expected to take several weeks (Reuters, Sep 10)

primary capacity, satellite imagery (Bloomberg, Aug 25)

every major Lukoil refinery is offline

Novorossiysk — main Black Sea port

fuel-oil terminal + the city, 4 killed (Sep 8–9)

crude outflow 800 → 350 kb/d, Jul → Aug — all three export directions now under attack

nationwide caps; Moscow 90% out of AI-92 (Euronews, Aug 20)

Lost production is also reducing government revenue

Russia's export restrictions leave less fuel available to other countries during the heating season. Sep 30 is the next deadline for the diesel export ban, and the jet-fuel export ban is scheduled to take effect on Nov 30. Damaged refineries also leave Russia with less fuel for its own gas stations and military as winter approaches.

gasoline & the remaining diesel

Imports aren't making up the difference. Fuel shipments on the Belarus rail route are running at 25 times last year's volume. The shortage also affects countries that rely on Russian fuel: Kyrgyzstan imports more than 90% of its gasoline from Russia and has about six weeks of reserves left.

Higher energy costs affect more than your fuel bill. They can slow business activity, keep inflation high, and eventually make food more expensive. People and businesses also use less oil when they can no longer afford it. Economists call this demand destruction. The charts below show that decline alongside recession estimates, interest rates, and the possible effects on food prices.

World petroleum & liquid fuels consumption, mb/d · monthly (EIA STEO, Sep 9 release) · Jan – Aug, 2026 vs 2025.

World oil use fell 4.3 million barrels per day below last year's level in May, a decline of about 4%. The gap narrowed to 3.6 million in July and 0.8 million in August. Using less oil helps contain prices, but it also reflects the strain on the economy. The IEA now expects demand to fall by 2.5 million barrels per day across 2026, compared with its August estimate of 1.6 million. It expects the quarterly decline to ease from 5.3 million in Q2 to 3.4 million in Q3 and 2.0 million in Q4, followed by a 2.6 million barrel per day recovery in 2027. The losses are concentrated in fuels such as diesel and in raw materials used to make chemicals, especially in Asia. These estimates aren't universally agreed on. OPEC expects demand to grow by 0.4 million barrels per day in 2026, a difference of 2.9 million between the two forecasts.

EIA STEO Table 3e (Sep 9 2026, forecast completed Sep 3) · Jan–Aug 2026 are actuals in that release · world/regional values are EIA estimates (apparent consumption, incl. refinery fuel & bunkering) · Sep 2026 onward is forecast, not shown.

Million b/d, July 2026 vs July 2025 — same table.

16.4 → 15.6 mb/d · part of Asia & Oceania; using stockpiles to support consumption

These estimates, published between June and September 2026, put the chance of a US recession over the next 12 months at 15% to 50%. Goldman Sachs has kept its estimate at 15% since Jun 26, down from 30% in late March. It repeated that estimate on Sep 14. Polymarket puts the odds at 32%, but covers a longer period, through the end of 2027. Keep that difference in mind when comparing the figures.

The Federal Reserve can cut interest rates to support a slowing economy, but persistent inflation makes that harder. August producer prices rose 5.4% from a year earlier. Consumer inflation held at 3.4%, with energy up 16.3%, while core inflation eased to 2.4%. Markets expect a rate increase at the Sep 16 meeting. Higher borrowing costs would add pressure as oil reserves are drawn down through the winter.

9-to-3 hold; officials 'see the need for a hike if inflation doesn't cool'

Odds of a September rate increase

futures markets, Sep 11 · up from about 72% Thursday after core CPI exceeded expectations · Polymarket: 62%

+0.4% for the month · July revised to 4.8% annually · energy +4.2%, diesel +24.1% annually · 10-year yield highest since Oct 2023

The European Central Bank says Germany and Italy could both be in a technical recession by the end of 2026 if the conflict continues.

You can see the wider effects in borrowing costs and everyday prices. Both are at multi-year highs. Together, they help explain why the Fed is considering higher interest rates even as the economy slows.

10-year US Treasury yield, % · chart through Sep 14 · weekly closes (FRED) · Sep 11 and 14: Yahoo closing values.

The 10-year Treasury yield is the rate the US government pays to borrow for a decade. It also influences mortgage and business loan rates. It has risen by about one percentage point, or 100 basis points, since before the war. Inflation and concerns about government debt are both putting pressure on rates. The national debt exceeds $40 trillion, and $8.4 trillion of Treasuries must be refinanced by year-end.

The 10-year yield closed at 4.975% on Sep 11 after briefly reaching 4.992%, its highest level since October 2023. It eased to 4.961% on Sep 14. Al Jazeera reported that it reached 5.02% during trading on Sep 15, its highest level since 2007, as traders anticipated a Federal Reserve rate increase. That latest quote isn't plotted; the chart shows closing values through Sep 14.

The 2-year yield was 4.63% on Sep 11, its highest since July 2024. The market's measure of expected inflation over the next 10 years eased to 2.36%. That suggests investors are seeking higher returns after inflation, even as their inflation expectations have fallen.

CPI (retail, amber) + PPI final demand (wholesale, blue) · % year-over-year · monthly (BLS) · Jan – Aug.

Consumer inflation rose from about 2.4% to 4.2% in three months as fuel became more expensive. It eased to 3.4% in July and stayed there in August. Prices rose 0.4% in August alone, with gasoline's 3.9% increase accounting for more than a third of that rise. Core inflation, which excludes food and energy, eased from 2.5% to 2.4% over the year, although its 0.3% monthly increase was above expectations. Producer prices rose faster, peaking at 5.9% in May and increasing 5.4% in August compared with a year earlier. July's reading was revised to 4.8%. Energy explains much of the increase, with diesel up 24.1%. These costs can reach businesses before they show up in household spending (BLS, Sep 10–11).

Natural gas is used to make ammonia, a key ingredient in nitrogen fertilizer. Higher gas and shipping costs can make food more expensive, but the effects may take more than a year to reach your grocery bill.

The timeline uses the 2007–08 and 2022 shocks to illustrate when those costs could reach food prices. It's a rough historical comparison, not a forecast from this model.

$4.34/bbl in Aug vs pre-war (Frontline)

Sep 14 · above $28/MMBtu Sep 10 (JOGMEC) · highest since Dec 2022

Sep 11 · high-$28s Sep 10 (JOGMEC) · highest in ~2.5 years

Higher energy prices can raise freight and fertilizer costs, affect planting and harvests, and eventually raise food prices. The bars show approximate windows based on the 2007–08 and 2022 shocks. Food prices could remain under pressure after oil reserves have reached the model's thresholds.

Upcoming reports and decisions that could change the outlook.

An official repair estimate for the East–West pipeline. AP reports 3–5 weeks. Analysts quoted by the Wall Street Journal estimate lost flow at more than 2.5 million barrels a day. Reuters reports that stocks at Yanbu, Saudi Arabia's Red Sea port, could support another five to seven days of exports.

The standoff scenario depends on this bypass. An official assessment that repairs will take only days would reverse the Sep 11 change in odds.

A new date for the Gulf–Iran talks in Salalah, postponed from Sep 14. Iran says Saudi Arabia requested the delay because of events in Yemen.

Resuming the talks could help restore tanker access through Hormuz.

Shipping conditions after the Houthis captured the Hanish islands on Sep 13–14. The Houthis claim 85 vessels passed through Bab el-Mandeb in 72 hours. Missile and drone attacks on Saudi cities wounded 13 civilians on Sep 13–14.

An attack on a non-Saudi vessel would raise the model's odds that the shipping corridor closes.

How banks respond to the Sep 14 sanctions on Russia's VTB; Treasury is meeting with financial institutions this week.

If banks stop handling VTB's payments, Iran loses channels for receiving oil revenue.

A verified count of vessels passing through Hormuz. Gen. Wright claims tankers are carrying 10 million barrels a day under Navy escort. Preliminary tracking shows four vessels on Sep 14, down from 14 a day a week earlier.

The claimed volume is roughly half the pre-war flow. The next count will help assess whether shipping activity supports that claim.

The Fed's rate decision — futures put a 25-basis-point increase at about 86% (Polymarket, 62%).

A hike would add borrowing-cost pressure on top of fuel prices.

The EIA's report for the week ending Sep 11. Watch for another week of slower reserve withdrawals and rising diesel stocks. The previous report showed withdrawals of 1.2 million barrels, down about 60% in a week. Diesel stocks stood at 106.3 million barrels.

The weekly figures help show whether supplies are tightening or recovering.

Russia's diesel export ban expires unless extended; the US-led coalition completes its withdrawal from Iraq; prediction-market bets settle.

Several deadlines in one week, each with supply or price implications.

The first EIA monthly outlook after the tanker attacks — its roughly $90 forecast for the second half is $16 below the latest Brent close.

Watch whether its view that shipping remains constrained but open survives.

The US midterm elections — President Trump has said the war will end just after the elections.

Watch whether fighting and diplomacy match the administration's stated timeline.

Russia's jet-fuel export ban takes effect.

Aviation fuel supplies tighten further as the world's remaining stocks run low.

These are estimates of where shortages could become more severe if the crisis continues. The first two dates use published stock levels; the others are inferred from customs and inventory data. Allow for uncertainty of a week or two.

US East Coast — Diesel and heating-oil stocks could fall below a month of supply. They are already 27% lower than last year.

Russia — The diesel export ban expires. With more than 30% of refining capacity damaged, Russia may have little fuel available to export.

China — Commercial oil stocks could begin to fall faster than normal.

Europe's oil hubs — Rotterdam-area diesel stocks could fall below 8.5–9M barrels, a level that would put pressure on trading. If the strait closes fully, the estimate moves up to mid-October.

Europe, at the pump — Shortages could reach consumers, with price increases putting pressure on governments.

Air travel — Russia's jet-fuel export ban begins Nov 30. The world's remaining stocks amount to about 26 days of flying.

Source: Hacker News

$ python3 numberwang.py 22
22... THAT'S NUMBERWANG!  (confidence: 99.3%)

$ python3 numberwang.py "45 - 44"
45 - 44... That's Wangernumb! Rotate the board!  (confidence: 100.0%)

$ python3 numberwang.py "hello how are you"
hello how are you... That's not even a number. It can never be Numberwang.  (confidence: 100.0%)

git clone https://github.com/GraafHenk/numberwang
cd numberwang
python3 numberwang.py 22

$ python3 numberwang.py
Welcome to Numberwang! (ctrl-c to stop playing Numberwang)
> zweiundzwanzig
zweiundzwanzig... THAT'S NUMBERWANG!  (confidence: 100.0%)
> shinty-six
shinty-six... That's not Numberwang.  (confidence: 100.0%)

from numberwang import load_model, wang_probabilities

model = load_model("model.json")
probs = wang_probabilities(model, "forty-seven")
# [p_not_numberwang, p_numberwang, p_not_a_number, p_wangernumb]

verdict = max(range(4), key=probs.__getitem__)

chars → Embedding(32) → Conv1d(128, k3) → ReLU
      → Conv1d(128, k3) → ReLU → global max pool
      → Linear(128) → ReLU → Linear(4) → softmax

pip install -r requirements.txt
python3 app.py

Source: Hacker News

TypeSafe announces System One models and Jev

TypeSafe announces System One models and Jev

Diogo Almeida, founder, TypeSafe

Models have been superhuman at chat for years, so where is all the automation?

This has been my driving question for the last four years. At OpenAI, I helped build the methods that made language models useful at following instructions and talking with people. That work ended up as the research behind ChatGPT.  At the time, I thought maybe chat models would lead to AGI, but despite the hype it became obvious to me that there was something really big missing.

After two years in stealth, countless technical challenges, and research breakthroughs… I am beyond excited to announce that today, TypeSafe AI is releasing our first System One Model: a new class of frontier models built to make fast, structured decisions that software can use directly.

We built a new stack entirely focused on automation: with a new model architecture, parallel sampler for maximum efficiency, and training method we call Reinforcement Learning for Calibrated Decisions (RLCD).

Our first public model is Jev, available today in early access. Jev achieves similar levels of intelligence on System One tasks compared to existing LLMs, while being two orders of magnitude faster and more efficient. While Jev gives up string generation, it’s optimized for structured outputs and can’t hallucinate. 

Think of Jev as a frontier-intelligence function call: unstructured state in, typed probabilistic decisions out. 

Extraordinary claims require extraordinary evidence so see below for the receipts. 💅

Reinforcement Learning with Human Feedback (RLHF) / Reinforcement Learning with Verifiable Rewards (RLVR)

Reinforcement Learning for Calibrated Decisions (RLCD)

Human preference: writeups and chat responses that human raters prefer.

Verifiable rewards: outputs that can be programmatically verified.

Calibrated decisions: answers with epistemically honest probabilities on System One tasks.

Unstructured data (e.g. text) with an emphasis on sequential messages.

Unstructured data (e.g. text) with an emphasis on structured program state.

Strings / generated text. Strings are flexible and can be anything: chat responses, code, hallucinations, refusals, or even type-safe structured values. To be used by software, responses need to be parsed + validated. There is also always some risk that the AI goes off the rails.

Type-safe structured values. Possible outputs and structure are defined in advance. The model never makes type errors. All answers are accompanied with calibrated probabilities and confidence scores.

Sequential. Generates one token at a time, each conditioned on the last.

Parallel. Generates all outputs in a single query. Incredibly efficient and hardware-aware.

Input tokens: from $0.20 to $10 / MTok.

Output tokens: ~5x more expensive than input tokens.

Input tokens: $0.042 / MTok ($42 per billion tokens).

Output tokens: FREE (too cheap to meter).

End-to-end response time is 3 to 329 seconds for frontier models.  Fast enough for interfacing with humans, but a big bottleneck when integrated in code.

End-to-end response time is 70ms-500ms for TypeSafe. This can range from 40x-200x faster for the same levels of frontier intelligence for System One shaped queries.

Even if prompted for a confidence estimate, models tend to be overconfident and inconsistent. If a model can do a task 95% of the time but doesn’t say when it’s in the 5%, it can’t automate that task.

Always communicates confidence and uncertainty with every output. Calibrated: higher confidence means higher accuracy. More consistent: returns similar answers for similar inputs.

Human-in-the-loop tasks (chatbots, copilots, coding agents). General and powerful, but requires human oversight because their freedom also means they might go off the rails.

Verifiable problems (math proofs, kernel optimization). When correctness can be checked cheaply and automatically, LLMs can generate, test, and iterate until they find something that works.

Demos. The flexibility of strings allows it to be incredible for quickly making prototypes that only work sometimes.

AI-Powered Workflows / smart if-statements. Structured outputs slot into ordinary software as fuzzy decision rules: classify, route, score, extract, or branch where hand-written logic is too brittle. The surrounding code constrains their freedom, making them easier to compose into reliable systems.

Map-reducing over big data. Turn petabytes of data into features and insights.Real-time applications. 100ms speeds means you can use AI in your applications where UX is critical.Verify everything. Score, judge, verify, guardrail, and detect jailbreaks of LLM prompts, reasoning traces, and/or outputs.

We love skeptics, and are skeptics ourselves.

There are some claims you can easily verify:

Speed per call: We truly are that fast, though our published evals are generally run from our laptops on the West Coast (this is where our service is currently based).

Cost per call: We make our pricing transparent. We can’t prove it isn’t subsidized; we’ll need the long-term to prove the sustainability of our pricing (which we expect to go down, not up).

No type errors: This would be an easy thing to falsify with just a single counter-example, but it is mathematically impossible.

For our bolder claims, we want to provide as much nuance as we can.

Our side-by-side demo shows a key difference between our models and LLMs: Jev outputs all probabilities in parallel instead of autoregressively generating by token. Strings are extremely powerful and general, but costly. “Giving up” strings actually gives us a lot of superpowers!

For people with early access to TypeSafe, here is the actual query.

The query is highly simplified and questions were chosen to have descriptive, human-readable keys so that the output on the screen is understandable.

The state is also a short, dense, and detailed paragraph, to emphasize the difference in sampling methodology. The relatively shorter input paints our model in an advantageous light.

For the keen eyed, for the recorded run, the only disagreement with GPT-5.6 Terra is on “Churn likelihood level”. The actual answer seems genuinely ambiguous to us.

We used GPT-5.6 Terra with default reasoning for this example, because we’ve found it to be the most comparable at intelligence to Jev on average.

Fun fact: a similar demo was what convinced us to go all-in in the direction of System One Models!

We made a new type of evaluation to measure how well AI works within code. We don’t optimize for a ground truth classification orand allow the harness and model to change (potentially allowing for overfitting via harness engineering). Instead, we assume there is a correct compute graph (a “workflow” represented in code) and use the predictions of the largest, smartest, and most expensive external models as reference probabilities.

Rephrased: every model gets the same workflow. We test how they compare to the average of the smartest models (in this case, Astra and Fable).

Jev is off the charts – owning the Pareto frontier for almost 2 orders of magnitude. We also compare to models with a generated prompt doing all the logic in their chain-of-thought, but this tends to do significantly worse than using the workflow itself.

Note that the calls here are significantly more complex than the side-by-side demonstration above. That’s because they’re more representative of the types of production workloads needed for true business automation. Below is the simplest of the 4 workflows we’re publishing:

The most reliable real-world workflows tend to have many independent, decomposed questions, with fine-grained behavior that’s dependent on probabilities instead of discrete decisions. The end result is discrete branching, but how we get to a final answer involves a lot of domain-specific engineering that needs to be done highly consistently.

See our workflow evals site for all the details: examples, disagreements, full queries, and each workflow.

This is where the claims of 193.6x faster, 444.6x cheaper on our home page comes from, and we expect that these are on the higher end of real world gains.

These content of these workflows were not deliberately chosen nor constructed to make our model look good, and are not in our training distribution. However, they were made by individuals on our model capabilities team, so some bias could exist.

We use the average of GPT-6 Astra and Fable 5.1 as the reference answer, which biases answers towards OpenAI and Anthropic’s models. We likely underestimate the relative performance of our model and DeepSeek’s models.

The LLMs use our System One LLM wrapper, which constrains LLMs to output structured decisions compatible with our API. We have found this to be the most accurate way to get decisions from LLMs, but this tends to be slower and more expensive than giving decisions without probabilities.

Hallucination and type-safety are intrinsically related, and we think the latter is table stakes for automation. Having a hallucinated tool call is inconvenient in an agent, but is an absolute deal-breaker if it’s part of a system with latency guarantees or it’s buried several layers deep in a dependency chain. Existing models, no matter how smart, still hallucinate and have type errors.

The numbers for LLMs are from OpenRouter i.e., there almost certainly is bias here: more complex queries might be routed to better models.

Our number is not empirical. Schema matching is guaranteed, thus we can confidently add 0% into the plots.

Perhaps the most exciting part of our work is enabling new use cases. We have a lot more to show you, but here are a couple of the team’s favorites:

We love how this doomo doomonstrates real-time intelligence and what can be doone with code + AI. The engineer behind it was worried about making 10 queries a second (which ends up costing ~$7/hour), but the rest of us agreed that was lower than expected! This is so fun we intend to not only release an in-depth walkthrough, but also host some events to hack on this.

The demo is on structured state as a data structure with text, not on images (yet…)

A non-AI doom bot could play better, but we wanted a bot that was reactive to different representations of game state, and most importantly… following instructions was cool as heck!

The objective of the game is to start on one Wikipedia page and reach a specific other Wikipedia page using only links you come across while traversing. Each step can mean choosing between hundreds to thousands of links! It’s a great playground for demonstrating not just intelligence-per-second, but also the compounding benefits of not hallucinating with high-cardinality choices.

As far as we know, it was completely random that both the 2nd and 3rd challenges started with “Rubber Duck.” The author only noticed when the team pointed it out.

Our speedups here tend to be a lot less than in previous demos. That’s because this is against the non-reasoning modes of the models (except Astra which was set to the lowest reasoning setting). This is also why Jev tended to finish in fewer steps (a sign of greater intelligence). This was to make the demo more bearable to watch. The LLMs look much worse at this task than with reasoning enabled.

Jev supports a cardinality up to 255. For the higher cardinality choices, we do a 2 stage-system of scoring independently then making an explicit choice, hence the occassional slowdown.

We’re still in Jev’s early days. We have a lot more in the pipeline and are so excited to keep on shipping 🔥.

Today, we are opening early access and bringing developers off the waitlist as quickly as we can. We want to hear which decisions you need to automate, where Jev works, and where it falls short. Tell us what sci-fi you want to build!!

We started TypeSafe because we believe that AI needs an interface software could depend on. We can't wait to see new use cases continuously diffuse through the community and economy.

Where do the names “System One Models” and “Jev” come from?

We were inspired by Daniel Kahneman, Thinking, Fast and Slow. The model class name draws on the distinction between fast, intuitive System 1 thinking and slow, deliberate System 2 reasoning.

“System 1 thinking” has also implied error-prone. For reasons we will get into in the future, we believe System One Models can be made more reliable than its alternatives.

We named Jev after William Stanley Jevons. We expect machine intelligence to follow a similar path to coal, after steam-engine efficiency led to an increase in demand. Every order of magnitude drop in the cost of intelligence unlocks orders of magnitude more use cases.

Why was a new training algorithm needed?

What use cases is Jev good for?

How does Jev perform against public benchmarks?

Where does our training data come from?

These are results are kinda crazy – how is it possible?

Source: Hacker News

Source: Hacker News