# doubleAI

> Source: https://aiwiki.ai/wiki/doubleai
> Updated: 2026-10-27
> Fact-checked: 2026-10-27
> Categories: AI Companies, AI Inference, AI Research, AI for Science
> License: CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/) - attribute to "AI Wiki (aiwiki.ai)"
> Cite as: AI Wiki. "doubleAI." aiwiki.ai, 27 Oct 2026. https://aiwiki.ai/wiki/doubleai
> From AI Wiki (https://aiwiki.ai), the free encyclopedia of artificial intelligence. Reuse freely with attribution.

**doubleAI** (legal name AA-I Technologies Ltd., also written AAI) is an Israeli [artificial intelligence](https://aiwiki.ai/wiki/artificial_intelligence) company founded in August 2023 by [Amnon Shashua](https://aiwiki.ai/wiki/amnon_shashua), Shai Shalev-Shwartz, Gal Beniamini, Yoav Levine, Or Sharir and Noam Wies.[1][2][3] It describes itself as a "deep frontier lab" focused on "the hardest problems in science and engineering."[1] The founders call their approach Artificial Expert Intelligence (AEI): systems meant to reach or exceed the level of top human experts in narrow technical fields, rather than broad general-purpose intelligence.[9][13] The company was registered in Jerusalem and later moved to offices in Ramat Gan, in the Tel Aviv area.[5][6]

doubleAI operated largely in stealth, announcing only what Shashua called a "partial unstealthing" in December 2024, and published research papers before it released any product.[4][8][9] Its $200 million Series A, co-led by [Lightspeed Venture Partners](https://aiwiki.ai/wiki/lightspeed_venture_partners) and [Bessemer Venture Partners](https://aiwiki.ai/wiki/bessemer_venture_partners) with participation from [Nvidia](https://aiwiki.ai/wiki/nvidia), Dell Technologies Capital and others, brought total funding to about $220 million.[13][26][30] doubleAI's announcements did not state a valuation; Calcalist reported that the round valued the company at more than $1 billion.[7][31] In March 2026 it released doubleGraph, an AI-generated rewrite of Nvidia's cuGraph library, and in October 2026 it opened early access to Maxwell, which it calls the first "code execution model": a system that rewrites a program's machine code while the program runs and formally proves each change correct.[13][15][26] The company's research agent, Double Agent, has also published results in graph theory, chess composition, quantum circuit design and hardware design.[19][20][22][25] As of October 2026 doubleAI had about 40 employees.[30][31][34]

## Overview

| Field | Detail |
| --- | --- |
| Legal name | AA-I Technologies Ltd.[2] |
| Brand | doubleAI (also reported as AAI or AA-I)[2][4][7] |
| Founded | August 2023, Jerusalem[3][5] |
| Headquarters | Ramat Gan, Israel (Tel Aviv area)[5][39] |
| Founders | Amnon Shashua, Shai Shalev-Shwartz, Gal Beniamini, Yoav Levine, Or Sharir, Noam Wies[1] |
| Chief executive | Amnon Shashua[1] |
| Chief technology officer | Shai Shalev-Shwartz[1] |
| Employees | About 40 (October 2026)[30][31][34] |
| Total funding | About $220 million (October 2026)[2][30] |
| Products and systems | Maxwell (code execution model, early access), Double Agent (research agent), doubleGraph (open-source cuGraph replacement)[15][25][26] |
| Website | doubleai.com[1] |

## History

### Origins and stealth period (2023-2024)

Shashua described the question behind the company in an August 2023 interview with Globes, given around the time doubleAI was founded: "Can AI imitate the qualities of a great mathematician or physicist? Can it invent the theory of relativity or develop the differential calculus as a tool to develop laws like those of Newton? It is a matter of abstraction that smart people know how to do and language models have not been trained for this. This is the next stage of AI."[3]

In a December 2024 post on X, Shashua traced the company to a lecture he had given at Reichman University about [large language models](https://aiwiki.ai/wiki/large_language_model). During the questions, mathematician Gil Kalai asked whether the technology could one day reach the level of a "great mathematician" or great scientist, and Shashua wrote that his "immediate instinct was negative." He and Shalev-Shwartz then set out to answer two questions: whether the trajectory of LLMs is subject to a performance ceiling, and if so, what leap would be needed to overcome it. Shashua wrote that three of his doctoral students (Levine, Sharir and Wies) graduated a few months later, and that the six of them (Shashua, Shalev-Shwartz, the three graduates and Gal Beniamini) founded AAI to pursue those questions.[8]

According to the Israeli companies registry, as reported by Globes, AA-I Technologies was registered in Jerusalem in August 2023.[3][5] Shashua told the Goldfarb Gross Seligman and PwC Israel "Up & Coming" conference in September 2024 that he had founded an AI company 10 months earlier and that it was "still in stealth," without naming it.[3] The company's existence became public that month when [Amazon Web Services](https://aiwiki.ai/wiki/amazon_web_services) listed AA-I among the 80 startups accepted into its 2024 Generative AI Accelerator. AWS described it only as a Jerusalem company that "develops an AI system that is capable of solving CS, math, and science problems at the highest level of human expertise."[3]

### Partial unstealthing (December 2024)

On December 3, 2024, Shashua announced what he called a "partial unstealthing" of the company, and its researchers released the paper "Artificial Expert Intelligence through PAC-reasoning," which set out the AEI framework.[8][9] Shashua wrote that the team was implementing the framework in "algorithmic expertise," using the competitive programming site Codeforces as a proving ground.[8] CTech described the announcement as the company proposing AEI as an alternative to the pursuit of artificial general intelligence, with PAC reasoning at its core.[37] Globes reported at the time that, according to industry estimates, the company had raised tens of millions of dollars.[4]

### Seed round, early research and Series A (2025)

In May 2025 Globes reported that the company had raised a seed round, which it then estimated at $20 million to $30 million, led by Dell Technologies Capital, the Barkat brothers' BRM Capital and Pitango First. The round had closed several months before the report, and the company then had several dozen employees, mainly AI engineers with doctorates.[5] Globes later put the seed round at $21 million.[6] Yair Snir, managing partner of Dell Technologies Capital, said in October 2026 that doubleAI was the first company developing its own models that the fund had backed.[31]

In July 2025 doubleAI published two research papers: "From Reasoning to Super-Intelligence: A Search-Theoretic Perspective" by Shalev-Shwartz and Shashua, and the [FormulaOne](https://aiwiki.ai/wiki/formulaone) benchmark paper, whose first author is Beniamini.[10][11] In August 2025 TIME included Shashua in its TIME100 AI list and described AAI as "a venture called AAI seeking to create superintelligent AI systems in STEM."[12]

On September 25, 2025, Calcalist reported that Lightspeed was leading a funding round worth hundreds of millions of dollars at a valuation of more than $1 billion.[7] On November 18, 2025, Globes reported that the $200 million round had been completed at a valuation of about $1 billion, that Nvidia had joined it, and that the company had raised $221 million in total.[6]

### WarpSpeed and doubleGraph (March 2026)

On March 2, 2026, doubleAI announced WarpSpeed, which it described as "the first Artificial Expert system to autonomously surpass world-class human experts in GPU performance engineering." It released doubleGraph, a rewritten version of Nvidia's cuGraph graph analytics library, on GitHub. The announcement also stated that doubleAI had recently raised a $200 million Series A co-led by Lightspeed and Bessemer.[13][15] The system was later presented under the name Maxwell, and the company's technical write-up of the cuGraph work now credits Maxwell.[14][2]

### Mobileye transition and the Maxwell launch (2026)

In July 2026 [Mobileye](https://aiwiki.ai/wiki/mobileye) said Shashua would step down as its chief executive once a successor was named.[30] CTech wrote that the change could give him more time for doubleAI, while noting that Shashua said his full attention remained on Mobileye.[33]

On October 16, 2026, speaking at the EY Journey conference, Shashua said the company had "decided to focus on code optimization" and would announce a product the following week. He also said doubleAI was exploring quantum computing and had moved into hardware design.[32] On October 19, 2026, the company introduced Maxwell and opened it for early access.[26] The Wall Street Journal reported the launch, and other outlets relayed its reporting.[2][30]

## Founders

| Name | Role at doubleAI | Background |
| --- | --- | --- |
| [Amnon Shashua](https://aiwiki.ai/wiki/amnon_shashua) | Co-founder and CEO[1] | Professor of computer science at the Hebrew University of Jerusalem; co-founder of [Mobileye](https://aiwiki.ai/wiki/mobileye), [AI21 Labs](https://aiwiki.ai/wiki/ai21_labs) and [Mentee Robotics](https://aiwiki.ai/wiki/mentee_robotics)[29][7] |
| Shai Shalev-Shwartz | Co-founder and CTO[1] | Professor of computer science at the Hebrew University and CTO of Mobileye; co-founder of Mentee Robotics[36][7] |
| Gal Beniamini | Co-founder[1] | PhD in theoretical computer science; background in systems and performance engineering; lead author of FormulaOne[29][16][11] |
| Yoav Levine | Co-founder[1] | Former doctoral student of Shashua; former AI21 Labs scientist; postdoctoral work at Stanford[8][5][6] |
| Or Sharir | Co-founder[1] | Former doctoral student of Shashua; former AI21 Labs scientist; postdoctoral work at Caltech[8][5][6] |
| Noam Wies | Co-founder[1] | Former doctoral student of Shashua; previously an algorithm specialist at OrCam[8][4] |

The company's first paper also listed Ido Ben-Shaul, Tomer Nussbaum and Shir Granot Peled as authors; Globes described Granot Peled as a former head of engineering at the blockchain company StarkWare.[9][5]

## Research program

### Artificial Expert Intelligence

doubleAI frames its work against both [artificial general intelligence](https://aiwiki.ai/wiki/artificial_general_intelligence) and narrow AI. The abstract of its first paper states that AEI "seeks to transcend the limitations of both Artificial General Intelligence (AGI) and narrow AI by integrating domain-specific expertise with critical, precise reasoning capabilities akin to those of top human experts."[9] In a March 2026 essay the company wrote that "humanity's progress is gated by experts" and argued that fields such as chip design, GPU performance engineering and molecular modeling advance "only as fast as the very small number of people who truly understand the details."[40] The company's website lists two core pillars: "search in the space of ideas," and verification, which it calls "the missing piece."[1]

Shashua has described the goal in terms of scientific ability. TIME reported that he is betting the lab may produce "great scientists at scale" within a few years. It also reported that Shashua believes, "for mathematical reasons," that aligning such superintelligent systems with human values and goals is impossible, quoting him: "I don't think this is a problem that will be solved. It's a problem that will be managed."[12]

### PAC reasoning and the Diligent Learner

The December 2024 paper, by Shalev-Shwartz, Shashua, Beniamini, Levine, Sharir, Wies, Ido Ben-Shaul, Tomer Nussbaum and Shir Granot Peled, introduced a framework it calls "Probably Approximately Correct (PAC) Reasoning," which it says provides "robust theoretical guarantees for reliably decomposing complex problems, with a practical mechanism for controlling reasoning precision." By analogy with System 1 (intuitive) and System 2 (reflective) thinking, the authors call this "System 3" for precise reasoning, inspired by the scientific method.[9]

The July 2025 paper by Shalev-Shwartz and Shashua argues that existing ways of learning from [chain-of-thought](https://aiwiki.ai/wiki/chain_of_thought) data, including supervised fine-tuning, [reinforcement learning](https://aiwiki.ai/wiki/reinforcement_learning), Tree-of-Thoughts and Monte Carlo Tree Search, suffer from distribution drift, a lack of embedded search and exponential inference costs. It proposes the "Diligent Learner," which models reasoning as a depth-first search guided by a validator, with backtracking on failure, and proves that under two assumptions it can learn efficiently from chain-of-thought data where the other methods cannot.[10] The company's March 2026 announcement named "Diligent Learning" and PAC reasoning and verification as the two ideas behind its GPU optimization results. It described the verification method as two components, one that searches for challenging test inputs and one that decides whether an output is correct, built on the premise that verification is simpler than search.[13]

In September 2026 the company published ArgMaxRL, an extension of the MaxRL reinforcement learning objective of Tajwar and colleagues from binary pass-or-fail rewards to graded rewards, aimed at tasks such as kernel optimization where correct answers differ in quality.[21]

### FormulaOne benchmark

FormulaOne, released in July 2025, is a benchmark of 120 dynamic programming problems on tree-like graphs, generated from monadic second-order logic, with a separate set of easier problems called FormulaOne-Warmup.[11][41] In the authors' evaluation, state-of-the-art models such as OpenAI's o3 solved less than 1% of the problems even when given 10 attempts and few-shot examples. The authors released the corpus together with an evaluation framework.[11]

## Maxwell

### Code execution models

doubleAI uses the term "code execution model" (CEM) for an AI system that works on a program while it runs, rather than writing source code. Its October 2026 announcement says a CEM's inputs are "the running program itself, the hardware it runs on, and the data flowing through it," and that it produces instructions meant only for the machine, while the source code "stays exactly as it was." Shashua said CEMs are "agents on behalf of the hardware as opposed to code agents that are on behalf of the human user."[26]

The company's manifesto describes Maxwell, formally "Maxwell's Daemon," as a system for superoptimization: finding the fastest way to carry out a piece of code on the machine in front of it. The name refers to the thought experiment James Clerk Maxwell proposed in 1867. According to the manifesto, Maxwell works much like a just-in-time compiler but may change the algorithm itself, and every replacement it makes must be accompanied by a formal proof that it is semantically equivalent to the original code. Faster versions may depend on conditions observed at runtime, and a guard hands control back to the original code if those conditions stop holding. The company says it has built a complete formalization of the Python language for this purpose and that Maxwell handles both CPU and GPU code, replacing Python with low-level code and [PyTorch](https://aiwiki.ai/wiki/pytorch) operations with kernels specialized for the GPU present.[27]

Co-founder Gal Beniamini told The Register that, unlike a compiler, Maxwell "can and does" change algorithms, and that its tracing and profiling add at most 2 percent overhead. The Register reported that developers install a runtime while the Maxwell model is hosted by doubleAI, an arrangement it noted "does require some amount of trust."[29] doubleAI uses a separate name, Double Agent, for the agent it applies to research problems such as mathematics and hardware design.[20][25]

### Reported results

All figures below are reported by doubleAI. CTech noted at launch that the headline figures had not been independently verified.[31]

| Date | Workload | Company-reported result |
| --- | --- | --- |
| March 2026 | Nvidia cuGraph graph analytics library (A100, L4 and A10G GPUs) | 3.6x mean speedup across every algorithm, 55% of algorithms above 2x and 18% above 10x; 192 specialized kernel sets per GPU architecture, 576 in total; released as doubleGraph[14][15] |
| June 2026 | Nvidia SOL-ExecBench, 235 Blackwell kernels | Beat Nvidia's optimized PyTorch baselines on 90% of kernels (2.24x on average) after one day of search, and on 233 of 235 (3.14x geometric mean) after two days[17] |
| July 2026 | [pandas](https://aiwiki.ai/wiki/pandas) on the 22-query PDS-H benchmark | 38.4x faster than stock pandas, with the same API and behavior[18] |
| October 2026 | LTX-2.5 video model, 4K generation on two [NVIDIA B200](https://aiwiki.ai/wiki/nvidia_b200) GPUs | 1.87x faster (134 seconds per clip versus 250 seconds)[28] |
| October 2026 | Demonstrations on the Maxwell website | [Whisper](https://aiwiki.ai/wiki/whisper) transcription on an RTX 4090 from 106.23 to 49.82 seconds; NetworkX betweenness centrality from 61.44 to 17.50 seconds[27] |

In the cuGraph write-up the company compared Maxwell with general [coding agents](https://aiwiki.ai/wiki/coding_agent). It reported that [Claude Code](https://aiwiki.ai/wiki/claude_code) and [OpenAI Codex](https://aiwiki.ai/wiki/openai_codex) produced correct implementations for 56% and 59% of the algorithms, and [Gemini CLI](https://aiwiki.ai/wiki/gemini_cli) for 22%, while Maxwell's output was correct in every case under doubleAI's own algorithm-specific verifiers. The same write-up says the work drew on "the trillion-parameter LRM we have trained," an agentic system based on "time travel," and an in-house environment scaled to thousands of GPUs.[14] It also describes reward hacking the team encountered, such as models that classified test inputs and returned precomputed answers instead of computing them.[14]

At launch the company said Maxwell had made data pipelines more than 38 times faster and AI inference kernels more than 4.5 times faster on real-world workloads; the Maxwell website listed the inference figure as 4.25 times.[26][27] The company told The Wall Street Journal that Maxwell cut one early customer's compute costs by 67%, and the Journal reported that LTX found Maxwell doubled the speed of its model inference.[30]

### Business model and availability

Maxwell opened for early access on October 19, 2026.[26] According to the Wall Street Journal's reporting, as relayed by The Next Web, doubleAI plans to charge customers a share of the compute costs they save, and Shashua expects AI labs with large compute bills, such as [OpenAI](https://aiwiki.ai/wiki/openai) and [Anthropic](https://aiwiki.ai/wiki/anthropic), to benefit most; he said even a 10% saving could be worth hundreds of millions of dollars to them.[30] CTech also reported the savings-based pricing plan.[31]

## Double Agent results

Double Agent is the name doubleAI uses for the AI agent it applies to expert-level tasks.[20] The company has reported the following results.

| Date | Field | Company-reported result |
| --- | --- | --- |
| August 2026 | Quantum cryptanalysis | Designed a point-addition quantum circuit for Shor's attack on ECDSA over the secp256k1 curve using 993,181 Toffoli gates on 1,205 qubits, which the company says is 2.5 times more efficient than Google's published designs and ahead of every submission on Eigen Labs' ecdsa.fail benchmark. The circuit was not published; doubleAI released a zero-knowledge proof of its existence and cost instead, and said the run cost under $5,000.[19] |
| August 2026 | Hardware design | An 82.5% reduction in cell count on the optimization tasks of Nvidia's CVDP hardware-design benchmark and 98.5% of evaluated problems solved, with each design checked against a formal contract derived from the specification[20] |
| October 2026 | Extremal graph theory | Explicit infinite families of d-regular graphs (d = q^3 + 1 for every prime power q) whose girth is at least 8/5 times the logarithm of the number of vertices to base d-1, up to a constant, improving the constant 4/3 from Margulis and from Lubotzky, Phillips and Sarnak that had been the best known since 1988; main theorem formalized in [Lean](https://aiwiki.ai/wiki/lean) on top of [Mathlib](https://aiwiki.ai/wiki/mathlib)[22][23] |
| October 2026 | Chess composition | A legal position in which all 63 of White's legal moves give check, beating a record of 60 that had stood since 1968, with a 190-ply game reaching it and a SAT-based proof, split into 7,224 cases, that 64 is impossible[25] |

The graph theory paper lists "Double Agent" as its sole author, which Shashua described as a paper written entirely by AI.[23][32] CTech reported that the company said the work cost less than $2,000 in computing and wrote that the result could be an early example of the kind of AI doubleAI is pursuing "if the result withstands scrutiny."[24] SiliconANGLE attributed the chess record to Maxwell, but doubleAI's own write-up credits Double Agent.[2][25]

## Funding

| Round | Reported or announced | Amount | Investors | Valuation |
| --- | --- | --- | --- | --- |
| Seed | Reported by Globes in May 2025 (closed several months earlier) | $21 million, per Globes (first reported as $20 million to $30 million) | Led by Dell Technologies Capital, BRM Capital and Pitango First[5][6] | Not disclosed |
| Series A | Reported as led by Lightspeed in September 2025; reported completed in November 2025; announced by the company on March 2, 2026 | $200 million | Co-led by Lightspeed Venture Partners and Bessemer Venture Partners, with Dell Technologies Capital, Greenfield Partners, Nvidia, Pitango and BRM; Globes also named Michael Dell among the participants[13][26][6] | Not disclosed by the company; Calcalist reported more than $1 billion and Globes about $1 billion[7][6][31] |

Neither report of the valuation specified whether the figure was pre-money or post-money.[6][7] The Wall Street Journal, as relayed by The Next Web, put total funding at $220 million; Globes put it at $221 million.[30][6] The company's website lists Lightspeed, Bessemer, Nvidia, Dell Technologies Capital, Pitango, Greenfield Partners and BRM Group as backers.[1]

## Compute

According to the Wall Street Journal, as relayed by The Next Web, doubleAI spent roughly half of its funding on 1,000 [Nvidia B300](https://aiwiki.ai/wiki/nvidia_blackwell_ultra) chips for its own use; SiliconANGLE described the same report as a plan to buy the chips.[30][2] Shashua told the Journal: "You cannot rely just on renting compute. Prices are going up." The Journal, citing [SemiAnalysis](https://aiwiki.ai/wiki/semianalysis), reported that the hourly price of [Nvidia H100](https://aiwiki.ai/wiki/nvidia_h100) chips on one-year contracts had risen 60% in a year.[30]

## Reception

CTech wrote that Maxwell's performance figures "have not been independently verified" and that the announcement did not give enough detail about tests, hardware configurations or baselines to judge how broadly the gains apply.[31] Startup Fortune noted that the graph theory result had not been peer reviewed and that a Lean proof confirms the internal logic but not whether the theorem is stated in the form the field cares about or is genuinely new.[35] TechRadar wrote that the company's stated goal to "copy and paste expertise into the world" raised "obvious questions about how far that idea can go in practice."[16]

In October 2026 LinkedIn ranked doubleAI second on its list of the top 15 startups in Israel, based on employee growth, jobseeker interest and talent attraction, and listed a company-provided headcount of 40.[34] Debarghya Das's "The Ultimate Neolab List" includes the company under the name "AAI (doubleAI)."[38]

## See also

- [List of neolabs](https://aiwiki.ai/wiki/list_of_neolabs)
- [Neolab](https://aiwiki.ai/wiki/neolab)
- [Amnon Shashua](https://aiwiki.ai/wiki/amnon_shashua)
- [FormulaOne](https://aiwiki.ai/wiki/formulaone)
- [Mobileye](https://aiwiki.ai/wiki/mobileye)
- [AI21 Labs](https://aiwiki.ai/wiki/ai21_labs)
- [Mentee Robotics](https://aiwiki.ai/wiki/mentee_robotics)
- [Automated theorem proving](https://aiwiki.ai/wiki/automated_theorem_proving)
- [KernelBench](https://aiwiki.ai/wiki/kernelbench)

## References

1. doubleAI. ["About doubleAI"](https://www.doubleai.com/about). doubleAI. Accessed October 27, 2026.
2. Deutscher, Maria. ["AI startup doubleAI debuts Maxwell 'code execution model'"](https://siliconangle.com/2026/10/19/ai-developer-doubleai-debuts-maxwell-code-execution-model/). SiliconANGLE, October 19, 2026.
3. Gilead, Assaf. ["Amnon Shashua's AI startup AA-I Technologies revealed"](https://en.globes.co.il/en/article-amnon-shashuas-ai-startup-aai-technologies-unveiled-1001489371). Globes, September 16, 2024.
4. Vaizberg, Meytal. ["Amnon Shashua's AA-I unveils new approach to AI"](https://en.globes.co.il/en/article-amnon-shashuas-aa-i-unveils-new-approach-to-ai-1001496057). Globes, December 4, 2024.
5. Gilead, Assaf. ["Shashua's artificial expert intelligence co AA-I raises funds"](https://en.globes.co.il/en/article-shashuas-artificial-expert-intelligence-co-aa-i-completes-seed-round-1001510682). Globes, May 18, 2025.
6. Gilead, Assaf. ["Nvidia joins $200m round raised by Shashua's AI startup"](https://en.globes.co.il/en/article-nvidia-joins-200m-round-raised-by-shashuas-ai-startup-1001526706). Globes, November 18, 2025.
7. Shulman, Sophie. ["Shashua's stealth AI startup raises hundreds of millions at over $1 billion valuation"](https://www.calcalistech.com/ctechnews/article/h1mz00im3xe). CTech (Calcalist), September 25, 2025.
8. Shashua, Amnon. ["The company AAI, founded a year ago, is still in stealth mode, but today we are doing a 'partial unstealthing'"](https://x.com/AmnonShashua/status/1863974561194103191). Post on X, December 3, 2024.
9. Shalev-Shwartz, Shai; Shashua, Amnon; Beniamini, Gal; Levine, Yoav; Sharir, Or; Wies, Noam; Ben-Shaul, Ido; Nussbaum, Tomer; Granot Peled, Shir. ["Artificial Expert Intelligence through PAC-reasoning"](https://arxiv.org/abs/2412.02441). arXiv:2412.02441, December 3, 2024.
10. Shalev-Shwartz, Shai; Shashua, Amnon. ["From Reasoning to Super-Intelligence: A Search-Theoretic Perspective"](https://arxiv.org/abs/2507.15865). arXiv:2507.15865, July 13, 2025 (revised July 26, 2025).
11. Beniamini, Gal, et al. ["FormulaOne: Measuring the Depth of Algorithmic Reasoning Beyond Competitive Programming"](https://arxiv.org/abs/2507.13337). arXiv:2507.13337, July 17, 2025.
12. Pillay, Tharin. ["Amnon Shashua"](https://time.com/collections/time100-ai-2025/7305878/amnon-shashua/). TIME100 AI 2025, TIME, August 2025.
13. doubleAI. ["doubleAI's WarpSpeed Beats a Decade of Expert-Engineered GPU Kernels: Every Single One of Them"](https://www.businesswire.com/news/home/20260302075567/en). Business Wire, March 2, 2026.
14. doubleAI. ["doubleAI's Maxwell: Surpassing Expert-Written Kernels At Scale"](https://www.doubleai.com/research/doubleais-maxwell-surpassing-expert-written-kernels-at-scale). doubleAI, March 5, 2026.
15. doubleAI. ["doubleGraph"](https://github.com/double-ai/doubleGraph). GitHub repository, created March 2026.
16. Athow, Desire. ["'AI can be relentless in a way that a human is not': With too few experts to go around, this startup is chasing a Matrix-style idea of 'copying and pasting' expertise"](https://www.techradar.com/pro/ai-can-be-relentless-in-a-way-that-a-human-is-not-with-too-few-experts-to-go-around-this-startup-is-chasing-a-matrix-style-idea-of-copying-and-pasting-expertise). TechRadar, March 19, 2026.
17. doubleAI. ["Maxwell approaches Speed of Light on Blackwell"](https://www.doubleai.com/research/maxwell-approaches-speed-of-light-on-blackwell). doubleAI, June 29, 2026.
18. doubleAI. ["Blazing Fast pandas Without Changing a Line"](https://www.doubleai.com/research/maxwell-pandas-blogpost). doubleAI, July 26, 2026.
19. doubleAI. ["Double Agent Discovers Record-Breaking ECDSA Cracking Circuit"](https://www.doubleai.com/research/double-agent-discovers-record-breaking-ecdsa-cracking-circuit). doubleAI, August 10, 2026.
20. doubleAI. ["Double Agent Brings Expert-Level AI to Hardware Design"](https://www.doubleai.com/research/double-agent-brings-expert-level-ai-to-hardware-design). doubleAI, August 25, 2026.
21. doubleAI. ["ArgMaxRL: Generalizing MaxRL to Continuous Rewards"](https://www.doubleai.com/research/argmaxrl-generalizing-maxrl-to-continuous-rewards). doubleAI, September 1, 2026.
22. doubleAI. ["Advancing a decades-old problem in extremal graph theory"](https://www.doubleai.com/research/advancing-a-decades-old-problem-in-extremal-graph-theory). doubleAI, October 6, 2026.
23. Double Agent (doubleAI). ["d-Regular Graphs of Girth (8/5) log(d-1) n"](https://www.doubleai.com/papers/girth85.pdf). doubleAI, October 2026.
24. CTech. ["OpenAI is solving decades-old math problems. An Israeli startup is taking a different route"](https://www.calcalistech.com/ctechnews/article/hswjujwr5). CTech (Calcalist), October 8, 2026.
25. doubleAI. ["A New Record in Extremal Chess Composition"](https://www.doubleai.com/research/maxwell-chess-record-63-checks). doubleAI, October 19, 2026.
26. doubleAI. ["doubleAI Unveils The First AI That Lives Inside Running Code"](https://www.financialcontent.com/article/bizwire-2026-10-19-doubleai-unveils-the-first-ai-that-lives-inside-running-code). Business Wire (via FinancialContent), October 19, 2026.
27. doubleAI. ["Maxwell: The first Code Execution Model"](https://heymaxwell.ai/manifesto). heymaxwell.ai, October 2026.
28. doubleAI. ["LTX-2.5 at 4K, in half the time"](https://www.doubleai.com/research/maxwell-ltx-2-5-4k-in-half-the-time). doubleAI, October 19, 2026.
29. Claburn, Thomas. ["Maxwell gives sub-optimal code a runtime glow-up"](https://www.theregister.com/ai-and-ml/2026/10/19/maxwell-gives-sub-optimal-code-a-runtime-glow-up/5303961). The Register, October 19, 2026.
30. Constantin, Ana Maria. ["Amnon Shashua's doubleAI launches an AI that speeds up code while it runs"](https://thenextweb.com/news/doubleai-maxwell-code-execution-model-compute-costs). The Next Web, October 19, 2026.
31. Shulman, Sophie. ["Shashua's doubleAI unveils AI models that optimize how software runs"](https://www.calcalistech.com/ctechnews/article/sk11xpc73ge). CTech (Calcalist), October 19, 2026.
32. Shulman, Sophie. ["Amnon Shashua's new AI venture is preparing to unveil its first product"](https://www.calcalistech.com/ctechnews/article/lojf2hi1m). CTech (Calcalist), October 16, 2026.
33. Orbach, Meir. ["From Mobileye to doubleAI and beyond, Amnon Shashua is preparing for his next act"](https://www.calcalistech.com/ctechnews/article/x3resmz2i). CTech (Calcalist), July 26, 2026.
34. LinkedIn News. ["LinkedIn Top Startups 2026: The 15 Israeli companies on the rise"](https://www.linkedin.com/pulse/linkedin-top-startups-2026-15-israeli-companies-rise-linkedin-news-ku2ef/). LinkedIn, October 14, 2026.
35. Fernandes, Elroy. ["An Israeli AI startup says it cracked a 38 year old graph theory bound for $2,000"](https://startupfortune.com/an-israeli-ai-startup-says-it-cracked-a-38-year-old-graph-theory-bound-for-2000/). Startup Fortune, October 15, 2026.
36. Shalev-Shwartz, Shai. ["Shai Shalev-Shwartz"](https://www.cs.huji.ac.il/~shais/). Hebrew University of Jerusalem, School of Computer Science and Engineering. Accessed October 27, 2026.
37. CTech. ["Amnon Shashua's new AI venture aims to push the limits of machine learning"](https://www.calcalistech.com/ctechnews/article/sywu003axyg). CTech (Calcalist), December 4, 2024.
38. Das, Debarghya. ["The Ultimate Neolab List"](https://debarghyadas.com/neolabs/). Accessed October 27, 2026.
39. doubleAI. ["Careers at doubleAI"](https://www.doubleai.com/careers). doubleAI. Accessed October 27, 2026.
40. doubleAI. ["Maxwell and the Need for Artificial Expert Intelligence"](https://www.doubleai.com/research/maxwell-and-the-need-for-artificial-expert-intelligence). doubleAI, March 12, 2026.
41. doubleAI. ["Research"](https://www.doubleai.com/research). doubleAI. Accessed October 27, 2026.