The Future of Life Institute (FLI) is a nonprofit organization which aims to steer transformative technology towards benefiting life and away from large-scale risks, with a focus on existential risk from advanced artificial intelligence (AI). FLI's work includes grantmaking, educational outreach, and advocacy within the United Nations, United States government, and European Union institutions. The founders of the Institute include MIT cosmologist Max Tegmark, UCSC cosmologist Anthony Aguirre, and Skype co-founder Jaan Tallinn. == Purpose == FLI's stated mission is to steer transformative technology towards benefiting life and away from large-scale risks. FLI's philosophy focuses on the potential risk to humanity from the development of human-level or superintelligent artificial general intelligence (AGI), but also works to mitigate risk from biotechnology, nuclear weapons and global warming. == History == === Founding === FLI was founded in March 2014 by MIT cosmologist Max Tegmark, Skype co-founder Jaan Tallinn, DeepMind research scientist Viktoriya Krakovna, Tufts University postdoctoral scholar Meia Chita-Tegmark, and UCSC physicist Anthony Aguirre. === Activism === Starting in 2017, FLI has offered an annual "Future of Life Award", with the first awardee being Vasili Arkhipov. The same year, FLI released Slaughterbots, a short arms-control advocacy film. FLI released a sequel in 2021. In 2018, FLI drafted a letter calling for "laws against lethal autonomous weapons". Signatories included Elon Musk, Demis Hassabis, Shane Legg, and Mustafa Suleyman. In January 2023, Swedish magazine Expo reported that the FLI had offered a grant of $100,000 to a foundation set up by Nya Dagbladet, a Swedish far-right online newspaper. In response, Tegmark said that the institute had only become aware of Nya Dagbladet's positions during due diligence processes a few months after the grant was initially offered, and that the grant had been immediately revoked. === Open letter on an AI pause === In March 2023, FLI published a letter titled "Pause Giant AI Experiments: An Open Letter". This called on major AI developers to agree on a verifiable six-month pause of any systems "more powerful than GPT-4" and to use that time to institute a framework for ensuring safety; or, failing that, for governments to step in with a moratorium. The letter said: "recent months have seen AI labs locked in an out-of-control race to develop and deploy ever more powerful digital minds that no-one - not even their creators - can understand, predict, or reliably control". The letter referred to the possibility of "a profound change in the history of life on Earth" as well as potential risks of AI-generated propaganda, loss of jobs, human obsolescence, and society-wide loss of control. Prominent signatories of the letter included Elon Musk, Steve Wozniak, Evan Sharp, Chris Larsen, and Gary Marcus; AI lab CEOs Connor Leahy and Emad Mostaque; politician Andrew Yang; deep-learning researcher Yoshua Bengio; and Yuval Noah Harari. Marcus stated "the letter isn't perfect, but the spirit is right." Mostaque stated, "I don't think a six month pause is the best idea or agree with everything but there are some interesting things in that letter." In contrast, Bengio explicitly endorsed the six-month pause in a press conference. Musk predicted that "Leading AGI developers will not heed this warning, but at least it was said." Some signatories, including Musk, said they were motivated by fears of existential risk from artificial general intelligence. Some of the other signatories, such as Marcus, instead said they signed out of concern about risks such as AI-generated propaganda. The authors of one of the papers cited in FLI's letter, "On the Dangers of Stochastic Parrots: Can Language Models Be Too Big?" including Emily M. Bender, Timnit Gebru, and Margaret Mitchell, criticised the letter. Mitchell said that “by treating a lot of questionable ideas as a given, the letter asserts a set of priorities and a narrative on AI that benefits the supporters of FLI. Ignoring active harms right now is a privilege that some of us don’t have.” === Open letter on prohibiting superintelligence === In October 2025, another letter, the "Statement on Superintelligence", was published. It called for a prohibition on the development of superintelligence not lifted before there is "broad scientific consensus that it will be done safely and controllably" and "strong public buy-in". FLI director Anthony Aguirre explained that "time is running out", expecting that the technology could arrive in as little as one to two years and counting on "widespread realization among society at all its levels" to stop it. He added that "whether it's soon or it takes a while, after we develop superintelligence, the machines are going to be in charge" and "that is not an experiment that we want to just run toward". The list of signatories included Nobel laureates Geoffrey Hinton, Daron Acemoglu, Beatrice Fihn, Frank Wilczek and John C. Mather as well as Hinton's fellow "godfather" of modern AI Yoshua Bengio, Steve Wozniak, Steve Bannon, Paolo Benanti, Prince Harry, Duke of Sussex and Meghan, Duchess of Sussex. The letter was also signed by the actors Joseph Gordon-Levitt and Stephen Fry, rapper Will.i.am and author Yuval Noah Harari. Former national security advisor Susan Rice, and OpenAI member of technical staff Leo Gao also signed their names to the letter. Polling released alongside the letter showed that 64% of American agreed that superintelligence "shouldn't be developed until it's provably safe and controllable" and only 5% believed it should be developed as quickly as possible. == Operations == === Advocacy === FLI has actively contributed to policymaking on AI. In October 2023, for example, U.S. Senate majority leader Chuck Schumer invited FLI to share its perspective on AI regulation with selected senators. In Europe, FLI successfully advocated for the inclusion of more general AI systems, such as GPT-4, in the EU's Artificial Intelligence Act. In military policy, FLI coordinated the support of the scientific community for the Treaty on the Prohibition of Nuclear Weapons. At the UN and elsewhere, the institute has also advocated for a treaty on autonomous weapons. === Research grants === The FLI research program started in 2015 with an initial donation of $10 million from Elon Musk. In this initial round, a total of $7 million was awarded to 37 research projects. In July 2021, FLI announced that it would launch a new $25 million grant program with funding from the Russian–Canadian programmer Vitalik Buterin. === Conferences === In 2014, the Future of Life Institute held its opening event at MIT: a panel discussion on "The Future of Technology: Benefits and Risks", moderated by Alan Alda. The panelists were synthetic biologist George Church, geneticist Ting Wu, economist Andrew McAfee, physicist and Nobel laureate Frank Wilczek and Skype co-founder Jaan Tallinn. Since 2015, FLI has organised biannual conferences with the stated purpose of bringing together AI researchers from academia and industry. As of April 2023, the following conferences have taken place: "The Future of AI: Opportunities and Challenges" conference in Puerto Rico (2015). The stated goal was to identify promising research directions that could help maximize the future benefits of AI. At the conference, FLI circulated an open letter on AI safety which was subsequently signed by Stephen Hawking, Elon Musk, and many artificial intelligence researchers. The Beneficial AI conference in Asilomar, California (2017), a private gathering of what The New York Times called "heavy hitters of A.I." (including Yann LeCun, Elon Musk, and Nick Bostrom). The institute released a set of principles for responsible AI development that came out of the discussion at the conference, signed by Yoshua Bengio, Yann LeCun, and many other AI researchers. These principles may have influenced the regulation of artificial intelligence and subsequent initiatives, such as the OECD Principles on Artificial Intelligence. The beneficial AGI conference in Puerto Rico (2019). The stated focus of the meeting was answering long-term questions with the goal of ensuring that artificial general intelligence is beneficial to humanity. == In the media == "The Fight to Define When AI is 'High-Risk'" in Wired. "Lethal Autonomous Weapons exist; They Must Be Banned" in IEEE Spectrum. "United States and Allies Protest U.N. Talks to Ban Nuclear Weapons" in The New York Times. "Is Artificial Intelligence a Threat?" in The Chronicle of Higher Education, including interviews with FLI founders Max Tegmark, Jaan Tallinn and Viktoriya Krakovna. "But What Would the End of Humanity Mean for Me?", an interview with Max Tegmark on the ideas behind FLI in The Atlantic.
Photo-consistency
In computer vision, photo-consistency determines whether a given voxel is occupied. A voxel is considered to be photo consistent when its color appears to be similar to all the cameras that can see it. Most voxel coloring or space carving techniques require using photo consistency as a check condition in Image-based modeling and rendering applications. == Usage == 3D Volumetric Reconstruction. Image registration. Multi-view reconstruction.
Digital cinema
Digital cinema is the digital technology used within the film industry to distribute or project motion pictures as opposed to the historical use of reels of motion picture film, such as 35 mm film. Whereas film reels have to be shipped to movie theaters, a digital movie can be distributed to cinemas in a number of ways: over the Internet or dedicated satellite links, or by sending hard drives or optical discs such as Blu-ray discs, then projected using a digital video projector instead of a film projector. Typically, digital movies are shot using digital movie cameras or in animation transferred from a file and are edited using a non-linear editing system (NLE). The NLE is often a video editing application installed in one or more computers that may be networked to access the original footage from a remote server, share or gain access to computing resources for rendering the final video, and allow several editors to work on the same timeline or project. Alternatively a digital movie could be a film reel that has been digitized using a motion picture film scanner and then restored, or, a digital movie could be recorded using a film recorder onto film stock for projection using a traditional film projector. Digital cinema is distinct from high-definition television and does not necessarily use traditional television or other traditional high-definition video standards, aspect ratios, or frame rates. In digital cinema, resolutions are represented by the horizontal pixel count, usually 2K (2048×1080 or 2.2 megapixels) or 4K (4096×2160 or 8.8 megapixels). The 2K and 4K resolutions used in digital cinema projection are often referred to as DCI 2K and DCI 4K. DCI stands for Digital Cinema Initiatives. As digital cinema technology improved in the early 2010s, most theaters across the world converted to digital video projection. Digital cinema technology has continued to develop over the years with RealD 3D, IMAX, RPX, 4DX, Dolby Cinema, and ScreenX, allowing moviegoers more immersive experiences. == History == The transition from film to digital video was preceded by cinema's transition from analog to digital audio, with the release of the Dolby Digital (AC-3) audio coding standard in 1991. Its main basis is the modified discrete cosine transform (MDCT), a lossy audio compression algorithm. It is a modification of the discrete cosine transform (DCT) algorithm, which was first proposed by Nasir Ahmed in 1972 and was originally intended for image compression. The DCT was adapted into the MDCT by J.P. Princen, A.W. Johnson and Alan B. Bradley at the University of Surrey in 1987, and then Dolby Laboratories adapted the MDCT algorithm along with perceptual coding principles to develop the AC-3 audio format for cinema needs. Cinema in the 1990s typically combined analog photochemical images with digital audio. Digital media playback of high-resolution 2K files has at least a 20-year history. Early video data storage units (RAIDs) fed custom frame buffer systems with large memories. In early digital video units, the content was usually restricted to several minutes of material. Transfer of content between remote locations was slow and had limited capacity. It was not until the late 1990s that feature-length films could be sent over the "wire" (Internet or dedicated fiber links). On October 23, 1998, Digital light processing (DLP) projector technology was publicly demonstrated with the release of The Last Broadcast, the first feature-length movie, shot, edited and distributed digitally. In conjunction with Texas Instruments, the movie was publicly demonstrated in five theaters across the United States (Philadelphia, Portland (Oregon), Minneapolis, Providence, and Orlando). === Foundations === In the United States, on June 18, 1999, Texas Instruments' DLP Cinema projector technology was publicly demonstrated on two screens in Los Angeles and New York for the release of Lucasfilm's Star Wars Episode I: The Phantom Menace. In Europe, on February 2, 2000, Texas Instruments' DLP Cinema projector technology was publicly demonstrated, by Philippe Binant, on one screen in Paris for the release of Toy Story 2. From 1997 to 2000, the JPEG 2000 image compression standard was developed by a Joint Photographic Experts Group (JPEG) committee chaired by Touradj Ebrahimi (later the JPEG president). In contrast to the original 1992 JPEG standard, which is a DCT-based lossy compression format for static digital images, JPEG 2000 is a discrete wavelet transform (DWT) based compression standard that could be adapted for motion imaging video compression with the Motion JPEG 2000 extension. JPEG 2000 technology was later selected as the video coding standard for digital cinema in 2004. In 1992, Hughes-JVC was founded by JVC and Hughes Electronics to develop ILA (Image Light Amplifer) digital video projectors for commercial movie theaters using liquid crystal on silicon (LCOS) technology. In 1997, JVC introduced D-ILA (Direct-Drive ILA) technology with a 2K resolution digital video projector. In 2000, JVC introduced a 4K resolution video projector using D-ILA technology. === Initiatives === On January 19, 2000, the Society of Motion Picture and Television Engineers, in the United States, initiated the first standards group dedicated to developing digital cinema. By December 2000, there were 15 digital cinema screens in the United States and Canada, 11 in Western Europe, 4 in Asia, and 1 in South America. Digital Cinema Initiatives (DCI) was formed in March 2002 as a joint project of many motion picture studios (Disney, Fox, MGM, Paramount, Sony Pictures, Universal and Warner Bros.) to develop a system specification for digital cinema. The same month it was reported that the number of cinemas equipped with digital projectors had increased to about 50 in the US and 30 more in the rest of the world. In April 2004, in collaboration with the American Society of Cinematographers, DCI created standard evaluation material (the ASC/DCI StEM material) for testing of 2K and 4K playback and compression technologies. DCI selected JPEG 2000 as the basis for the compression in the system the same year. Initial tests with JPEG 2000 produced bit rates of around 75–125 Mbit/s for 2K resolution and 100–200 Mbit/s for 4K resolution. === Worldwide deployment === In China, in June 2005, an e-cinema system called "dMs" was established and was used in over 15,000 screens spread across China's 30 provinces. DMs estimated that the system would expand to 40,000 screens in 2009. In 2005, the UK Film Council Digital Screen Network launched in the UK by Arts Alliance Media creating a chain of 250 2K digital cinema systems. The roll-out was completed in 2006. This was the first mass roll-out in Europe. AccessIT/Christie Digital also started a roll-out in the United States and Canada. By mid-2006, about 400 theaters were equipped with 2K digital projectors with the number increasing every month. In August 2006, the Malayalam digital movie Moonnamathoral, produced by Benzy Martin, was distributed via satellite to cinemas, thus becoming the first Indian digital cinema. This was done by Emil and Eric Digital Films, a company based at Thrissur using the end-to-end digital cinema system developed by Singapore-based DG2L Technologies. In January 2007, Guru became the first Indian film mastered in the DCI-compliant JPEG 2000 Interop format and also the first Indian film to be previewed digitally, internationally, at the Elgin Winter Garden in Toronto. This film was digitally mastered at Real Image Media Technologies in India. In 2007, the UK became home to Europe's first DCI-compliant fully digital multiplex cinemas; Odeon Hatfield and Odeon Surrey Quays (in London), with a total of 18 digital screens, were launched on 9 February 2007. By March 2007, with the release of Disney's Meet the Robinsons, about 600 screens had been equipped with digital projectors. In June 2007, Arts Alliance Media announced the first European commercial digital cinema Virtual Print Fee (VPF) agreements (with 20th Century Fox and Universal Pictures). In March 2009, AMC Theatres announced that it closed a $315 million deal with Sony to replace all of its movie projectors with 4K HDR digital projectors starting in the second quarter of 2009; it was anticipated that this replacement would be finished by 2012. As digital cinema technology improved in the early 2010s, most theaters across the world converted to digital video projection. In January 2011, the total number of digital screens worldwide was 36,242, up from 16,339 at end 2009 or a growth rate of 121.8 percent during the year. There were 10,083 d-screens in Europe as a whole (28.2 percent of global figure), 16,522 in the United States and Canada (46.2 percent of global figure) and 7,703 in Asia (21.6 percent of global figure). Worldwide progress was slower as in some territories, particularly Latin America and Africa. As of 31 March 2015, 38,719 screens (out of a total of 3
Anti-social Media Bill (Nigeria)
Anti-social Media Bill was introduced by the Senate of the Federal Republic of Nigeria on 5 November 2019 to criminalise the use of the social media in peddling false or malicious information. The original title of the bill is Protection from Internet Falsehood and Manipulations Bill 2019. It was sponsored by Senator Mohammed Sani Musa from the largely conservative northern Nigeria. After the bill passed second reading on the floor of the Nigeria Senate and its details were made public, information emerged on the social media accusing the sponsor of the bill of plagiarising a similar law in Singapore which is at the bottom of global ranking in the freedom of speech and of the press. But the senator denied that he plagiarised Singaporean law. == Opposition to the bill == Angry reactions trailed the introduction of the bill, and a number of civil society organisations, human rights activists, and Nigerian citizens unanimously opposed the bill. International rights group, Amnesty International and Human Rights Watch condemned the proposed legislation saying it is aimed at gagging freedom of speech which is a universal right in a country of over two hundred million people. Opposition political parties are very critical of the bill and accused the government of attempting to strip bare, Nigerian citizens of their rights to free speech and destroying same social media on whose power and influence the ruling All Progressives Congress, APC came to power in 2015. Nigeria Information Minister, Lai Mohammed has been at the center of public criticism because he is suspected to be the brain behind the proposed act. Lai was a former spokesman of then opposition All Progressives Congress. A "Stop the Social Media Bill! You can no longer take our rights from us" online petition campaign to force the Nigeria parliament to drop the bill received over 90,000 signatures within 24 hours. In November 2019, after the bill passed second reading in the senate, Akon Eyakenyi, a senator from Akwa Ibom State publicly said he would resist the bill. === Support for the bill === Those who support the proposed act especially Senators have often argued that the law would help curtail hate speech. President Muhammad Buhari who is seen as a beneficiary of the influence and power of the social media and free speech has been mute about it. But the president's senior aides and family members have publicly spoken in support of the bill. In November 2019, the wife of the president, Aisha Buhari, told a gathering at the Nigeria's National Mosque in the capital, Abuja that if China with over one billion people could regulate the social media, Nigeria should do same. But Nigerians reacted saying Nigeria is not a one-party communist state like China. Days later, a daughter to the president, Zahra Indimi told a gathering of young people in Abuja that social media had become a potent weapon for bullying those they thought were doing better than them in terms of social class and called for a critical regulation. == Key provisions of the bill == === Title === Protection from Internet Falsehoods, Manipulations and Other Related Matters Bill 2019. === Explanatory memorandum === This Act is to prevent Falsehoods and Manipulations in Internet transmission and correspondences in Nigeria. To suppress falsehoods and manipulations and counter the effects of such communications and transmissions and to sanction offenders with a view to encouraging and enhancing transparency by Social Media Platforms using the internet correspondences. === Objectives === One objective of the bill is to prevent the transmission of false statements or declaration of facts in Nigeria. Another objective of the bill is to end the financing of online mediums that transmit false statements. Measures will be taken to detect and control inauthentic behaviour and misuse of online accounts (parody accounts). When paid content is posted towards a political end, there will be measures to ensure the poster discloses such information. There will be sanction for offenders. === Transmission of false statement === According to the bill, a person must not: Transmit a statement that is false or, Transmit a statement that might: i. Affect the security or any part of Nigeria. ii. Affect public health, public safety or public finance. iii. Affect Nigeria's relationship with other countries. iv. influence the outcome of an election to any office in a general election. v. Cause enmity or hatred towards a person or group of persons. Anyone guilty of the above is liable to a fine of N300,000 or three years' imprisonment or both (for individual); and a fine not exceeding ten million naira (for corporate organisations). Same punishment applies for fake online accounts that transmit statements listed above. === Parody accounts === The bill says a person shall not open an account to transmit false statement. Anyone found guilty will be fined N200,000 or three years' imprisonment or both (for an individual) or five million naira (for corporate organisations). If such accounts transmit a statement that will affect security or influence the outcome of an election, such a person will be fined N300,000 or three years' imprisonment or both. If a person receives payment or reward to help another to transmit false statements knowingly, he/she is liable to a fine of N150,000 or three years' imprisonment or both. If a person receives payment or reward to help another to transmit a statement affects security or influence the outcome of an election, the fine is N300,000 or three years' imprisonment or both (for individual) and ten million naira for organisations. === Declaration === According to the bill, a law enforcement department can issue a "declaration" to offenders. And this declaration will be issued even if the "false statement" has been corrected or pulled down. The offender will be required to publish a "correction notice" in a specified newspaper, online location or other printed publication of Nigeria. Failure to comply, a person is liable to N200,000 or 12 months' imprisonment or both (for individual) and five million naira for organisations. === Access blocking order === The bill says the law enforcement department will also issue an access blocking order to offenders. The law enforcement department may direct the NCC to order the internet access service provider to disable access by users in Nigeria to the online location and the NCC must give the internet access service provider an access blocking order. An internet access service provider that does not comply with any access blocking order is liable on conviction to a fine not exceeding ten million naira for each day during any part of which that order is not fully complied with, up to a total of five million naira.
Europa-Lehrmittel
The Verlag Europa-Lehrmittel Nourney, Vollmer GmbH & Co. KG (English: European educational media), based in Haan (district Gruiten) near Düsseldorf, Germany, develops educational media for vocational training and further education as used in vocational schools, universities, companies and further education institutions for industrial-technical, business, social, health as well as gastronomic professions. The contents are delivered in print as well as in digital formats, including offered as apps, simulations, software and through an exam preparation portal. The publisher has its own digital learning platform, the Europathek. General partner is Nourney, Vollmer & Co. GmbH, a school and specialist book publisher. == History == The publishing house was founded in Wuppertal in 1948, when the first books were published. In Eislingen/Fils a construction office was set up to create the technical drawings for the textbooks - today the drawing office of the publisher in Ostfildern. A little over 40 years later, in 1989, a new company building was built in Haan-Gruiten. In 1997, Europa-Lehrmittel acquired the Fachbuchverlag Pfanneberg with its program aimed at gastronomic professions. Further program expansions in vocational training were realized by adding titles of the publishing houses Lau, Gerber, Pluspunkt, parts of the Verlag Handwerk und Technik program and that of the Gildebuchverlag. In 2013, the scientific and technical university titles of Verlag Harri Deutsch were taken over, including the bestselling "Handbook of Mathematics" by Ilya Nikolaevich Bronstein and Konstantin Adolfovic Semendyayev. In addition, Europa-Lehrmittel's first exam apps appeared. In 2017, the publisher introduced its own digital learning platform named Europathek. The Prüfungsdoc (English: Examination Doc) platform, which enables targeted online preparation for exams, went online the same year. In 2018, the titles of the Düsseldorf publishing house SOL for self-organized learning were incorporated into Europa-Lehrmittel's program. == Program == The program meanwhile includes well over 2000 print and digital publications, including specialist titles on metal technology, automotive and electrical engineering (like "Tabellenbuch Metall" (English: Metal Book of Tables)), on business administration, as well as works for the gastronomic profession (e.g. "Der junge Koch" (English: The Young Cook)). Many titles also appear as licensed editions in over 20 different languages on all continents. All contents are presented in an annual catalog, on the website, in schools, in companies, and at trade fairs and congresses. The publisher is a member of the Deutscher Hauswirtschaftsrat. == Digital educational media == Europathek: In addition to digital books, Europathek also provides media packages, additional materials and e-learning contents. The online contents of the 'media shelf' can be used per web browser. There are also software versions and apps for offline use, with the help of which the obtained titles can be viewed after downloading. Prüfungsdoc (English: Examination Doc): Prüfungsdoc offers web-based learning, practice and repetition to prepare for the intermediate and final examinations. This includes simulations of the exams with direct evaluation of the results.
PagedAttention
PagedAttention is an attention algorithm for efficient serving of large language models (LLMs). It was introduced in 2023 by Woosuk Kwon and colleagues in the paper Efficient Memory Management for Large Language Model Serving with PagedAttention, alongside the vLLM serving engine. The method stores the key–value cache used during autoregressive decoding in fixed-size blocks that can be mapped to non-contiguous physical memory, borrowing ideas from virtual memory, paging, and operating system design. == Background == In transformer inference, the key–value cache grows with sequence length and the number of concurrent requests. Kwon et al. argued that earlier serving systems typically reserved contiguous cache regions in advance, which caused reserved space, internal fragmentation, and external fragmentation. In their experiments, the paper reported that the effective memory utilization of previous systems could fall as low as 20.4%. == Description == PagedAttention partitions the cache of each sequence into fixed-size KV blocks. A request's cache is represented as a sequence of logical blocks, while a block table maps those logical blocks to physical GPU-memory blocks. As a result, neighboring logical blocks do not need to be contiguous in physical memory, and new blocks can be allocated on demand as generation proceeds. The design also makes it easier to share cache state across related decoding paths. In vLLM, physical blocks can be reference-counted and shared among requests or branches, with block-granularity copy-on-write used when a shared block must be modified. The original paper applied this design to parallel sampling, beam search, and prompts with shared prefixes. == Mathematical formulation == For a query token i {\displaystyle i} in causal self-attention, the standard attention output can be written as a i j = exp ( q i ⊤ k j / d ) ∑ t = 1 i exp ( q i ⊤ k t / d ) , o i = ∑ j = 1 i a i j v j {\displaystyle a_{ij}={\frac {\exp(\mathbf {q} _{i}^{\top }\mathbf {k} _{j}/{\sqrt {d}})}{\sum _{t=1}^{i}\exp(\mathbf {q} _{i}^{\top }\mathbf {k} _{t}/{\sqrt {d}})}},\;\mathbf {o} _{i}=\sum _{j=1}^{i}a_{ij}\mathbf {v} _{j}} where q i {\displaystyle \mathbf {q} _{i}} , k j {\displaystyle \mathbf {k} _{j}} , and v j {\displaystyle \mathbf {v} _{j}} are the query, key, and value vectors, and d {\displaystyle d} is the attention dimension. If the cache is partitioned into blocks of size B {\displaystyle B} , the key and value blocks may be written as K j = ( k ( j − 1 ) B + 1 , … , k j B ) , V j = ( v ( j − 1 ) B + 1 , … , v j B ) {\displaystyle \mathbf {K} _{j}=(\mathbf {k} _{(j-1)B+1},\ldots ,\mathbf {k} _{jB}),\;\mathbf {V} _{j}=(\mathbf {v} _{(j-1)B+1},\ldots ,\mathbf {v} _{jB})} PagedAttention then performs the computation blockwise: A i j = exp ( q i ⊤ K j / d ) ∑ t = 1 ⌈ i / B ⌉ exp ( q i ⊤ K t / d ) , o i = ∑ j = 1 ⌈ i / B ⌉ V j A i j ⊤ {\displaystyle \mathbf {A} _{ij}={\frac {\exp(\mathbf {q} _{i}^{\top }\mathbf {K} _{j}/{\sqrt {d}})}{\sum _{t=1}^{\lceil i/B\rceil }\exp(\mathbf {q} _{i}^{\top }\mathbf {K} _{t}/{\sqrt {d}})}},\;\mathbf {o} _{i}=\sum _{j=1}^{\lceil i/B\rceil }\mathbf {V} _{j}\mathbf {A} _{ij}^{\top }} where A i j {\displaystyle \mathbf {A} _{ij}} is the vector of attention scores for the j {\displaystyle j} -th KV block. In the formulation given by Kwon et al., this preserves the causal attention calculation while allowing the key and value blocks to reside in non-contiguous physical memory. == Performance and use == The vLLM paper reported that, on its evaluated workloads, the use of PagedAttention and the associated memory-management design improved serving throughput by 2–4× over the compared baselines, including FasterTransformer and Orca, while preserving model outputs. In experiments on OPT-13B with the Alpaca trace, the paper also reported memory savings of 6.1–9.8% for parallel sampling and 37.6–55.2% for beam search through KV-block sharing. A 2024 survey of LLM serving systems described PagedAttention as having become an industry norm in LLM serving frameworks, citing support in TGI, vLLM, and TensorRT-LLM. == Limitations and alternatives == Subsequent work has described trade-offs in the approach. The 2025 vAttention paper argued that PagedAttention requires attention kernels to be rewritten to support paging and increases software complexity, portability issues, redundancy, and execution overhead, proposing instead a memory manager that keeps the cache contiguous in virtual memory while relying on demand paging for physical allocation. === vAttention === Unlike PagedAttention, vAttention does not introduce a different attention rule; it retains the standard attention computation Attention ( q i , K , V ) = softmax ( q i K ⊤ s c a l e ) V . {\displaystyle \operatorname {Attention} (q_{i},K,V)=\operatorname {softmax} \left({\frac {q_{i}K^{\top }}{\mathrm {scale} }}\right)V.} In the notation of Prabhu et al., the key and value tensors for a request seen so far are K , V ∈ R L ′ × ( H × D ) {\displaystyle K,V\in \mathbb {R} ^{L'\times (H\times D)}} , where L ′ {\displaystyle L'} is the context length seen so far, H {\displaystyle H} is the number of KV heads on a worker, and D {\displaystyle D} is the dimension of each KV head. In systems prior to PagedAttention, the K cache (or V cache) at each layer of a worker is typically allocated as a 4D tensor of shape [ B , L , H , D ] , {\displaystyle [B,L,H,D],} where B {\displaystyle B} is batch size and L {\displaystyle L} is the maximum context length supported by the model. vAttention preserves this contiguous virtual-memory view while deferring physical-memory allocation to runtime. A serving framework maintains separate K and V tensors for each layer, so vAttention reserves 2 N {\displaystyle 2N} virtual-memory buffers on a worker, where N {\displaystyle N} is the number of layers managed by that worker. The maximum size of one virtual-memory buffer is B S = B × S , {\displaystyle BS=B\times S,} where S {\displaystyle S} is the maximum size of a single request's per-layer K cache (or V cache) on a worker. The paper defines S = L × H × D × P , {\displaystyle S=L\times H\times D\times P,} where P {\displaystyle P} is the number of bytes needed to store one element. In this formulation, vAttention keeps the KV cache contiguous in virtual memory and relies on demand paging for physical allocation, rather than modifying the attention kernel to operate over non-contiguous KV-cache blocks.
Bare machine
In information technology, a bare machine (or bare-metal computer) is a computer which has no operating system. The software executed by a bare machine, commonly called a bare metal program or bare metal application, is designed to interact directly with hardware. Bare machines are widely used in embedded systems, particularly in cases where resources are limited or high performance is required. == Bare machine computing == Bare Machine Computing is a computing paradigm in which application software runs directly on a bare machine as a single, stand-alone executable, without an operating system or device drivers. The application software has direct access to hardware resources, and there is typically no distinction between user and kernel mode. It is self-managed software that boots, loads and runs without using any other software components. Bare metal programs are typically written in a close-to-hardware language such as C or assembly language. == Advantages == Typically, a bare-metal application will run faster, use less memory and be more power efficient than an equivalent program that relies on an operating system, due to the inherent overhead imposed by system calls. For example, hardware inputs and outputs are directly accessible to bare metal software, whereas they must usually be accessed through system calls when using an OS. It has no OS and therefore has no OS-related vulnerabilities. == Disadvantages == Bare metal applications typically require more effort to develop because operating system services such as memory management and task scheduling are not available. Debugging a bare-metal program may be complicated by factors such as: Lack of a standard output. The target machine may differ from the hardware used for program development (e.g., emulator, simulator). This forces setting up a way to load the bare-metal program onto the target (flashing), start the program execution and access the target resources. == Examples == === Early computers === Early computers, such as the PDP-11, allowed programmers to load a program, supplied in machine code, to RAM. The resulting operation of the program could be monitored by lights, and output derived from magnetic tape, print devices, or storage. Amdahl UTS's performance improves by 25% when run on bare metal without VM, the company said in 1986. === Embedded systems === Bare machine programming is a common practice in embedded systems, in which microcontrollers or microprocessors boot directly into monolithic, single-purpose software without loading an operating system. Such embedded software can vary in structure. For example, one such program paradigm, known as foreground-background or superloop architecture, consists of an infinite main loop in which each task is executed sequentially and must voluntarily return control back to the loop. The loop runs these cooperative background processes that are not time-critical, while interrupt service routines momentarily interrupt the loop to handle time-critical foreground tasks.