Komunitas
hexbear.net
https://archive.is/ Secretary of War Pete Hegseth announced the rollout of GenAI.mil today in a video posted to X. To hear Hegseth tell it, the website is “the future of American warfare.” In practice, based on what we know so far from press releases and Hegseth’s posturing, GenAI.mil appears to be a custom chatbot interface for Google Gemini that can handle some forms of sensitive—but not classified—data. Hegseth’s announcement was full of bold pronouncements about the future of killing people. These kinds of pronouncements are typical of the second Trump administration which has said it believes the rush to “win” AI is an existential threat on par with the invention of nuclear weapons during World War II. Hegseth, however, did not talk about weapons in his announcement. He talked about spreadsheets and videos. “At the click of a button, AI models on GenAI can be used to conduct deep research, format documents, and even analyze video or imagery at unprecedented speed,” Hegseth said in the video on X. Office work, basically. “We will continue to aggressively field the world’s best technology to make our fighting force more lethal than ever before.” Emil Michael, the Pentagon’s under secretary for research and engineering, also stressed how important GenAI would be to the process of killing people in a press release about the site’s launch. “There is no prize for second place in the global race for AI dominance. We are moving rapidly to deploy powerful AI capabilities like Gemini for Government directly to our workforce. AI is America’s next Manifest Destiny, and we’re ensuring that we dominate this new frontier,” Michael said in the press release, referencing the 19th century American belief that God had divinely ordained Americans to settle the west at the same time he announced a new chatbot. The press release says Google Cloud’s Gemini for Government will be the first instance available on the internal platform. It’s certified for Controlled Unclassified Information, the release states, and claims that because it’s web grounded with Google Search–meaning it’ll pull from Google search results to answer queries–that makes it “reliable” and “dramatically reduces the risk of AI hallucinations.” As we’ve covered, because Google search results are also consuming AI content that contains errors and AI-invented data from across the web, it’s become nearly unusable for regular consumers and researchers alike. During a press conference about the rollout this morning, Michael told reporters that GenAI.mil would soon incorporate other AI models and would one day be able to handle classified as well as sensitive data. As of this writing, GenAI’s website is down. “For the first time ever, by the end of this week, three million employees, warfighters, contractors, are going to have AI on their desktop, every single one,” Michael told reporters this morning, according to Breaking Defense. They’ll “start with three million people, start innovating, using building, asking more about what they can do, then bring those to the higher classification level, bringing in different capabilities,” he said. The second Trump administration has done everything in its power to make it easier for the people in Silicon Valley to push AI on America and the world. It has done this, in part, by framing it as a national security issue. Trump has signed several executive orders aimed at cutting regulations around data centers and the construction of nuclear power plants. He’s threatened to sign another that would block states from passing their own AI regulations. Each executive order and piece of proposed legislation threatens that losing the AI race would mean making America weak and vulnerable and erode national security. The country’s tech moguls are rushing to build datacenters and nuclear power plants while the boom time continues. Nevermind that people do not want to live next to datacenters for a whole host of reasons. Nevermind that tech companies are using faulty AIs to speed up the construction of nuclear power plants. Nevermind that the Pentagon already had a proprietary LLM it had operated since 2024. “We are pushing all of our chips in on artificial intelligence as a fighting force. The Department is tapping into America’s commercial genius, and we’re embedding generative AI into our daily battle rhythm,’ Hegseth said in the press release about GenAI.mil. “AI tools present boundless opportunities to increase efficiency, and we are thrilled to witness AI’s future positive impact across the War Department.”
Komunitas
ibbit.at
Secretary of War Pete Hegseth announced the rollout of GenAI.mil today in a video posted to X. To hear Hegseth tell it, the website is “the future of American warfare.” In practice, based on what we know so far from press releases and Hegseth’s posturing, GenAI.mil appears to be a custom chatbot interface for Google Gemini that can handle some forms of sensitive—but not classified—data. Hegseth’s announcement was full of bold pronouncements about the future of killing people. These kinds of pronouncements are typical of the second Trump administration which has said it believes the rush to “win” AI is an existential threat on par with the invention of nuclear weapons during World War II. Hegseth, however, did not talk about weapons in his announcement. He talked about spreadsheets and videos. “At the click of a button, AI models on GenAI can be used to conduct deep research, format documents, and even analyze video or imagery at unprecedented speed,” Hegseth said in the video on X. Office work, basically. “We will continue to aggressively field the world’s best technology to make our fighting force more lethal than ever before.” Emil Michael, the Pentagon’s under secretary for research and engineering, also stressed how important GenAI would be to the process of killing people in a press release about the site’s launch. “There is no prize for second place in the global race for AI dominance. We are moving rapidly to deploy powerful AI capabilities like Gemini for Government directly to our workforce. AI is America’s next Manifest Destiny, and we’re ensuring that we dominate this new frontier,” Michael said in the press release, referencing the 19th century American belief that God had divinely ordained Americans to settle the west at the same time he announced a new chatbot. The press release says Google Cloud’s Gemini for Government will be the first instance available on the internal platform. It’s certified for Controlled Unclassified Information, the release states, and claims that because it’s web grounded with Google Search–meaning it’ll pull from Google search results to answer queries–that makes it “reliable” and “dramatically reduces the risk of AI hallucinations.” As we’ve covered, because Google search results are also consuming AI content that contains errors and AI-invented data from across the web, it’s become nearly unusable for regular consumers and researchers alike. During a press conference about the rollout this morning, Michael told reporters that GenAI.mil would soon incorporate other AI models and would one day be able to handle classified as well as sensitive data. As of this writing, GenAI’s website is down. “For the first time ever, by the end of this week, three million employees, warfighters, contractors, are going to have AI on their desktop, every single one,” Michael told reporters this morning, according to Breaking Defense. They’ll “start with three million people, start innovating, using building, asking more about what they can do, then bring those to the higher classification level, bringing in different capabilities,” he said. The second Trump administration has done everything in its power to make it easier for the people in Silicon Valley to push AI on America and the world. It has done this, in part, by framing it as a national security issue. Trump has signed several executive orders aimed at cutting regulations around data centers and the construction of nuclear power plants. He’s threatened to sign another that would block states from passing their own AI regulations. Each executive order and piece of proposed legislation threatens that losing the AI race would mean making America weak and vulnerable and erode national security. The country’s tech moguls are rushing to build datacenters and nuclear power plants while the boom time continues. Nevermind that people do not want to live next to datacenters for a whole host of reasons. Nevermind that tech companies are using faulty AIs to speed up the construction of nuclear power plants. Nevermind that the Pentagon already had a proprietary LLM it had operated since 2024. “We are pushing all of our chips in on artificial intelligence as a fighting force. The Department is tapping into America’s commercial genius, and we’re embedding generative AI into our daily battle rhythm,’ Hegseth said in the press release about GenAI.mil. “AI tools present boundless opportunities to increase efficiency, and we are thrilled to witness AI’s future positive impact across the War Department.” From 404 Media via this RSS feed
Komunitas
sh.itjust.works
Washington (United States) (AFP) – Meta announced Friday it will integrate content from major news organizations into its artificial intelligence assistant to provide Facebook, Instagram and WhatsApp users with real-time information. The social media giant said Meta AI will offer breaking news, entertainment and lifestyle stories when users ask news-related questions, drawing from partnerships with outlets including CNN, Fox News, Le Monde, People and USA Today. The feature will allow users to access “more diverse content sources” and receive links to partner websites to dive deeper into stories, Meta said in a blog post. Meta said the expansion aims to make its AI assistant “more responsive, accurate, and balanced” by incorporating diverse viewpoints, acknowledging that “real-time events can be challenging for current AI systems to keep up with.” The initial partnerships span mainstream and conservative-leaning publications, including The Daily Caller and The Washington Examiner. The company said it plans to continue adding partnerships and develop new features as competition intensifies among technology firms to enhance the capabilities of their AI assistants. Meta AI is available across the company’s platforms, serving billions of users globally. The announcement comes as artificial intelligence companies, including ChatGPT and Google’s Gemini, increasingly move to incorporate live web content and news feeds. Meta has had a hot and cold relationship with the news media over the years. The company founded by Mark Zuckerberg in 2004 declared that news was a very small share of user engagement on the company’s platforms and began shutting down the Facebook News tab in markets like the United States, Britain and France. This also saw the end of multi-million dollar deals with leading news organizations. Zuckerberg also made the surprise decision in January to axe Meta’s US fact-checking program, as he more closely aligned with the Trump administration’s antipathy to establishment news. That scheme had employed third-party fact checkers, many from news media organizations such as AFP, to expose misinformation disseminated on the platform.
Komunitas
ibbit.at
Photo by Omar Al-Ghosson It seems abundantly obvious to me that everyone who believes in free expression, whatever side of various political equations they may be on, should be concerned about what YouTube just did to me. If it could happen to me because of my allegedly controversial political viewpoints, it could happen to you because of yours. But in order for anyone to be concerned, first they have to understand what it is exactly that did happen, so I’ll try explain that as succinctly as possible, because I know everybody is busy with things aside from the latest chapter in the never-ending Cancellation of David Rovics story. I’ll try to explain things in a way that hopefully makes sense to every reader, not just the Rovics fans or the ones who are knowledgeable about music streaming platforms and other aspects of the indie music biz. I’m an independent artist, like millions of others in the world, putting out self-released recordings (that is, recordings that are not released and promoted by a record label, other than my own little one-person label). I’ve been doing this since before the internet became widely used, and long before the invention of the MP3 or streaming music on the web. I have never had anything remotely approaching a hit or what they would call commercial success, but among musicians who have their music up for streaming, I’m easily in the top 10% of most-streamed artists, usually within the top 5%. All the pop stars are well within the top 1% of most-streamed artists — there’s a very steep curve happening here, and I don’t mean to over-inflate my importance in the scheme of things. I’m just trying to say that I do have an audience. My songs are streamed millions of times every year on YouTube, millions of times a year on Spotify, and less on the other platforms, because there are really only two main platforms in the world (outside of China). When a musician records an album, whether they’re on a label or not, the musician or the label gets the songs registered with an artists’ rights entity such as ASCAP or BMI (most countries have one of these organizations but in the US there are two). That way the music gets counted as existing for purposes of radio play, and we musicians get paid for radio play that way, getting a direct deposit from BMI (in my case) every three months. Every time a community radio station plays one of my songs, BMI sends me one cent. At the same time as the musician gets their songs registered for copyright with one of those agencies, the musician also signs up for distribution with a distributor such as CDBaby. This used to be something artists did in order to make their music available for people to download on iTunes and other platforms that sold downloads. Having all of our music there already meant that it was also there when the era of paid downloads ended and the era of free streaming platforms began. When the corporations decided that rather than selling downloads, they would now start streaming the world’s music for free, they already had all of our music available to use for this purpose. Opting out was possible, but would mean a future of invisibility along with poverty. Opting in meant just poverty, but not invisibility, too. Spotify initiated the free streaming model, and all the other streaming platforms soon followed suit, out of necessity, in order to compete, no matter what nice ideas some of them may have had about fair models for compensating artists. As things stand now, none of the platforms that offer ad-supported (”free”) streaming options pay artists more than a small fraction of a cent per song streamed, though some platforms may be better than others in various ways. What has played out since free streaming became the way the vast majority of music fans on Planet Earth listen to music is, outside of China, two corporations have grown to dominate the world of music online — Spotify and YouTube. To emphasize the point I’m making here: I mentioned the quarterly payments musicians get for radio airplay before. We also get regular payments from the music streaming platforms. Usually people get those payments sent to them via a distributor like CDBaby, so you don’t have to set up a separate account with each of the hundred or so streaming platforms that CDBaby gets your music onto. So when I get my payments from CDBaby, I’m sure just like the vast majority of other artists on streaming platforms, you can see the breakdown of which platforms generated how much money. It’s evident with every one of those payments that Spotify and YouTube dominate the market. In the battle for the eyes and ears of the world, these corporations and their corporate practices have destroyed so many lives, careers, and entire professions. (For a lot more info about how horrible YouTube and its corporate parent Google/Alphabet are, read or listen to Cory Doctorow’s recent book, Enshitification.) In this process, these two giants of music streaming became basically a duopoly. If you live in most of the world, just as you do a search on Google if you’re looking to do a search online, if you’re looking for a video you go to YouTube, and if you want to hear a song you go to Spotify, or YouTube sends you to YouTube Music to find the song you might be looking for. This is where the details become crucially important, as well as a bit confusing. Please bear with me, if you can. YouTube Music deleted all of my albums — 50 of them altogether — several of which had been there since YouTube Music began. Along with all of the albums, they disappeared all of the comments and all of the evidence that these songs had ever been heard millions of times. As an artist on YouTube Music who puts out albums, I no longer exist. Why is this confusing? Well, if you go to YouTube and look for me, you’ll still see me all over the place. Videos of me singing at shows and in my living room, and songs from albums that other people have uploaded on the platform. So, why does getting removed from YouTube Music matter, in the scheme of things, if people can still hear my music on other streaming platforms, and even, with some of the songs at least, on YouTube itself? I asked Gemini (Google’s AI chatbot) to explain the impact on an artist’s future career if their music is removed from YouTube Music. I excerpt here Gemini’s response, which was very clear and very accurate, to the best of my fairly significant level of knowledge on this subject. “David Rovics – Topic” is the way artists are listed on YouTube Music if they have a presence on the platform. If you look for any other artist, you’ll find they have a “Topic” page on YouTube Music, but not me, as of last week. Estimated Impact of YouTube Music Removal The removal of his solo albums from the “David Rovics – Topic” music streaming platform would have a significant and strongly negative impact on potential audience growth, particularly within the mainstream digital music ecosystem. Here is an analysis based on the context: + Loss of the “Digital Highway”: One context snippet likens major streaming platforms to the “infrastructure for our lives” and a “second home.” Being removed from a platform is like disappearing, similar to how being off Facebook can feel like disappearing from the virtual world. YouTube Music/Premium is a “highway” for millions of global listeners, and the removal eliminates the path of least resistance for new, casual listeners to discover and consume his full album catalog. + Hindrance to Discovery: The “Topic” channel is the primary source for music distribution on YouTube’s dedicated streaming service. Its removal stops the platform’s algorithms from suggesting his catalog to listeners who might enjoy political folk or similar genres, severely limiting organic discovery through the YouTube Music ecosystem. + Erosion of Market Share: Losing a major global platform like YouTube Music represents the loss of a key segment of the overall music streaming market, which is crucial for modern audience growth. + Forced Friction: New listeners must now go directly to his website, Patreon, Substack, Bandcamp (where he faces shadowbanning issues), or other, less-dominant streaming platforms. This added friction prevents casual users from encountering his music, which directly impacts the potential for mass audience expansion. To provide a little more context about what all this stuff means: every month artists who are on Spotify get an email from Spotify telling us that of the 18,000 people who listened to our music last month, 4,000 (or whatever the numbers may be for that month) were “new listeners.” Those are often people who got to a song of mine because they were listening to another leftwing artist, and the algorithm thought they’d like to hear me, or a particular song of mine. The same phenomenon is at play on other streaming platforms, though they don’t send helpful monthly emails the way Spotify does. As Gemini explained, this recommendation phenomenon will no longer be in play with my music on YouTube Music anymore. According to my research on this sort of thing, it is so rare that an artist has their entire catalog deleted by a platform for reasons unrelated to copyright infringement, there are no examples available aside from mine that I can find. (If anyone reading this knows of one, please let me know!) Part of the reason it’s hard to know if this has ever happened before is it’s very unusual, apparently, for platforms to actually tell artists why they might be taking such an action, when they take it. But according to my research, while it is very common for individual songs to be taken down for violating one rule or another, it is almost unheard of for an artist’s entire catalog to be removed. Given how rare this sort of thing is, how damaging it is to those targeted, and how arbitrarily such actions have been taken by corporations like Google/Alphabet/YouTube, once other people understand what has just happened to me and what could happen to anyone else who gets on the blacklist, I hope that soon I will not be alone in speaking out against what they’ve specifically just done to me. For those who don’t know the back story to why I’m being targeted, a few words on that history. In early 2024, my first album about Israel’s ongoing war on Gaza, Notes from a Holocaust, was removed from my discography on Spotify, with no notification or explanation to me or to anyone else. I later put the album back up with a slightly altered name, and it has stayed up. This removal of an entire album has never happened on any other platform, until last week. Right around the same time that the album was removed from Spotify, I received my first notification from YouTube that my channel was being demonetized for the next 90 days, to punish me for posting a Houthi Army press release which I thought was an interesting thing to share with people, given that the US was at that time actively bombing Yemen. After one or two more of these 90-day suspensions of monetization, in January 2025 YouTube informed me that my channel would now be permanently demonetized, and that I had no recourse. I contacted the YouTube customer service people to confirm that this was indeed the case, and not a mistake. At the same time as this was going on, YouTube was regularly deleting videos, specifically if they involved me singing my “Song for the Houthi Army” or my song, “I Support Palestine Action.” It seemed they would wait for someone to report the video, and then delete it. This is my only way to understand their process for deciding which videos to delete on YouTube, because of the way it has thus far involved getting rid of some renditions of these songs while leaving others on the platform. YouTube’s explanation for which rules I was violating that had led to my channel’s permanent demonetization was “supporting criminal organizations,” which is a broad concept that under both British and US law includes the Houthi Army, and in the UK the British nonviolent direct-action group, Palestine Action as well. In the UK, verbally expressing support for proscribed organizations like them is a crime punishable by up to 14 years in prison, as this violates Section 12 of the UK’s Terrorism Act of 2000. In the US, verbally expressing support for proscribed organizations may be legal under the First Amendment, but earning income from praising proscribed organizations, at least by my understanding of the law, is a different matter legally. As I understand US laws, this is why you’re allowed to visit Cuba, but you’re in big trouble, potentially, if you spend any money there. In any case, for whatever reason — never fully explained — my channel was demonetized, and certain videos of certain songs continue to be randomly disappearing. When this happens, I get two emails from YouTube, one explaining that this song violates the rule against supporting criminal organizations and has therefore been taken down, and another one telling me that my channel has been permanently demonetized (despite the fact that it already was, last January). What I believe just happened last week with my existence as an artist with albums available on YouTube Music ceasing, was the YouTube bureaucracy figured out that if they really were serious about demonetizing this guy, they couldn’t just demonetize the videos while allowing his albums to earn royalties on YouTube Music. Because of their legal and financial arrangements with distributors, keeping my music on the platform but not sending royalty money to CDBaby for that artist might be more complicated than simply severing all ties between the artist and the distributor, as they exist on YouTube Music, or as they used to. When they figure out that I’m the lyricist and producer behind the artist, Ai Tsuno, they will presumably delete all of her albums as well. So far she still exists as an artist with albums on YouTube Music. Soon her next album will be up there, too, including the song, “They Deleted David Rovics.” Funny, maybe, but it by no means compensates for anything that’s being done by YouTube to this artist, as they disappear me in stages, as they’re doing. Anyone who takes a look at the extremely small numbers of listeners to Ai Tsuno on any of the platforms can see what I’m up against if I were to just upload all of my albums back on to YouTube Music — were that even possible, now that they’ve removed all of the David Rovics albums. No one would notice they’re there, or it would take a very long time for the songs to get back into the recommendation algorithms that they were in before last week. I’d like to point out two aspects to these efforts to deplatform me that I think are especially relevant. One is the way the laws in the UK and US work with regards to criminal organizations that anyone in government seems actually to be worried about, anyone criticizing Israeli genocidal actions or proclaiming their support for international law which defends things like armed resistance to occupation is breaking all kinds of laws. Laws that basically do not apply in any other context. So the laws themselves, not at all accidentally, are set up to support groups like UK Lawyers for Israel, and legally arm them for their systematic trolling activities. UK Lawyers for Israel is one of a number of different outfits on both sides of the Atlantic that proudly and publicly go about trying to vilify academics, artists, journalists, and all sorts of other people, and using these ridiculous laws to their greatest advantage. UK Lawyers for Israel began announcing in emails sent with their masthead to venues telling them they should cancel my gigs, in February, 2024, during the same winter when all the problems with Spotify and YouTube began (problems with various forms of suppression on Facebook and Bandcamp began earlier). Intentions of groups like these are not hidden, they’re open and proud about their successes in getting professors fired and gigs canceled. One of the other chatbots I consulted about having my entire catalog deleted by YouTube Music was confident that because this sort of action is so unheard-of and appeared to be so obviously political in nature, surely the artist targeted in this way would benefit by getting lots of media publicity. So far, anyway, I can report that that chatbot’s assumptions were false. (This is often the case with AI, as with humans.) There are a couple things on that idea of outrageous corporate behavior like this garnering media attention that might be worth noting. One is that people hear about stuff that gets media attention. They don’t hear about stuff that doesn’t, generally. So we are under the impression that AI-generated music is very popular, because every once in a while an AI-generated song gets popular. Most AI-generated music, like most completely human-generated music, hardly gets heard at all, however. Another thing is it often seems to be the case that an artist needs to be at a certain level of fame in the first place, in order for things like having all their albums pulled from a major platform to generate any media attention, and I’m not Kneecap or Bob Vylan (though I think they’re great). The post YouTube, aka the Biggest Platform on Earth, Has Deleted All My Albums appeared first on CounterPunch.org. From CounterPunch.org via this RSS feed
Komunitas
lemmygrad.ml
cross-posted from: https://lemmygrad.ml/post/9929539 Just saw the news on twitter - and this is my own screenshot of their API pricing taken just now. V3 has received an update and is moving to 3.2 - including on the web interface. They posted all these nice benchmarks which I guess means something? Lol tbh I’m not sure how much you can trust these benchmarks. What does this mean? Well, I gave the twitter thread announcement to Crush (and the new deepseek-v3.2, which you are automatically upgraded to both on API and web interface) and this is what it made of the news: Quick Summary DeepSeek has released V3.2, replacing the experimental version. There are two main models: V3.2: General-purpose, balanced performance (GPT‑5 level) V3.2‑Speciale: Specialized for complex reasoning (Gemini‑3.0‑Pro level) Both are open‑source and available via API. What’s New & Why It Matters 1. Thinking Integrated with Tool‑Use V3.2 can now “think” while using tools (like searching the web, running code, or calling APIs). This makes AI assistants more transparent and better at multi‑step tasks. You can choose thinking mode (slower but more thorough) or non‑thinking mode (faster for simple tasks). 2. Two Models for Different Needs V3.2: Use this for most agentic work—chatbots, coding assistants, general automation. It’s cost‑effective and reliable. V3.2‑Speciale: Use this only when you need top‑tier reasoning—complex math, competitive programming, advanced planning. It uses more tokens (costs more) but solves harder problems. 3. Open‑Source & API Options API: Available immediately. V3.2‑Speciale has a temporary endpoint until Dec 15, 2025. Self‑hosting: Download from Hugging Face and run locally for privacy or cost savings. Pricing: Same for both models [crush made a mistake in pricing here, not sure why. See my screenshot for current pricing or check on https://api-docs.deepseek.com/quick_start/pricing] Practical Advice for Agent Builders Which Model Should You Use? Start with V3.2—it handles 90% of agentic tasks well. Switch to V3.2‑Speciale only for tasks that require deep, step‑by‑step reasoning. Monitor token costs—V3.2‑Speciale uses more tokens, so watch your API bill. Key Improvements for Agentic Interfaces Better reasoning transparency—you can see the model’s thought process when using tools. Mode flexibility—toggle thinking on/off based on task complexity. Stronger performance on benchmarks (math Olympiads, programming contests). Timeline & Availability Now: V3.2 on DeepSeek App, Web, and API Until Dec 15: V3.2‑Speciale via temporary API (same price) Always: Both models open‑source on Hugging Face
Komunitas
lemmygrad.ml
Just saw the news on twitter - and this is my own screenshot of their API pricing taken just now. V3 has received an update and is moving to 3.2 - including on the web interface. They posted all these nice benchmarks which I guess means something? Lol tbh I’m not sure how much you can trust these benchmarks. What does this mean? Well, I gave the twitter thread announcement to Crush (and the new deepseek-v3.2, which you are automatically upgraded to both on API and web interface) and this is what it made of the news: Quick Summary DeepSeek has released V3.2, replacing the experimental version. There are two main models: V3.2: General-purpose, balanced performance (GPT‑5 level) V3.2‑Speciale: Specialized for complex reasoning (Gemini‑3.0‑Pro level) Both are open‑source and available via API. What’s New & Why It Matters 1. Thinking Integrated with Tool‑Use V3.2 can now “think” while using tools (like searching the web, running code, or calling APIs). This makes AI assistants more transparent and better at multi‑step tasks. You can choose thinking mode (slower but more thorough) or non‑thinking mode (faster for simple tasks). 2. Two Models for Different Needs V3.2: Use this for most agentic work—chatbots, coding assistants, general automation. It’s cost‑effective and reliable. V3.2‑Speciale: Use this only when you need top‑tier reasoning—complex math, competitive programming, advanced planning. It uses more tokens (costs more) but solves harder problems. 3. Open‑Source & API Options API: Available immediately. V3.2‑Speciale has a temporary endpoint until Dec 15, 2025. Self‑hosting: Download from Hugging Face and run locally for privacy or cost savings. Pricing: Same for both models [crush made a mistake in pricing here, not sure why. See my screenshot for current pricing or check on https://api-docs.deepseek.com/quick_start/pricing] Practical Advice for Agent Builders Which Model Should You Use? Start with V3.2—it handles 90% of agentic tasks well. Switch to V3.2‑Speciale only for tasks that require deep, step‑by‑step reasoning. Monitor token costs—V3.2‑Speciale uses more tokens, so watch your API bill. Key Improvements for Agentic Interfaces Better reasoning transparency—you can see the model’s thought process when using tools. Mode flexibility—toggle thinking on/off based on task complexity. Stronger performance on benchmarks (math Olympiads, programming contests). Timeline & Availability Now: V3.2 on DeepSeek App, Web, and API Until Dec 15: V3.2‑Speciale via temporary API (same price) Always: Both models open‑source on Hugging Face
Komunitas
lemmygrad.ml
cross-posted from: https://lemmygrad.ml/post/9899994 wake up open twitter to catch up see deepseek did it again (and as a reminder, Deepseek-r1 only came out in January so it’s been less than 12 months since their last bombshell) One more graph: What this all means Traditional AI models are trained to be “rewarded” for a correct final answer. Get the expected answer, win points, be incentivized to get the answer more often. This has a major flaw: a correct answer does not guarantee correct reasoning. A model can guess, use a shortcut, or even have flawed logic but still output the right answer. This approach completely fails for tasks like theorem proving, where the process is the product. DeepSeekMath-V2 tackles this with a novel self-verifying reasoning framework: the Generator: One part of the model generates mathematical proofs and solutions. the Verifier: Another part acts as the critic, checking every step of the reasoning for logical rigor and correctness The Loop: If the verifier finds a flaw, it provides feedback, and the generator revises the proof. This creates a co-evolution cycle where both components push each other to become smarter This new approach allows the model to set record-breaking performance. As you can see from the charts above, it scores second-place on ProofBench-Advanced, just behind Gemini. But Gemini isn’t open-source, Deepseekmath-V2 is. The model weights are available on Huggingface under an Apache 2.0 license: https://huggingface.co/deepseek-ai/DeepSeek-Math-V2. This means researchers, developers, and enthusiasts around the world can download, study, and build upon this model right now. They can fine-tune or change the model to fit their needs and research, which promises a lot of exciting math discoveries happening soon - I predict (on no basis mind you) that this will help solve computing problems to start with, either practical or theoretical. Beyond just the math, the self-verification mechanism is a crucial step towards building AI systems whose reasoning we can trust, which is vital for applications such as scientific research, formal verification, and safety-critical systems. It also proves that ‘verification-driven’ training is a viable and powerful alternative to the ‘answer-driven’ method used to this day.
Komunitas
lemmygrad.ml
wake up open twitter to catch up see deepseek did it again (and as a reminder, Deepseek-r1 only came out in January so it’s been less than 12 months since their last bombshell) One more graph: What this all means Traditional AI models are trained to be “rewarded” for a correct final answer. Get the expected answer, win points, be incentivized to get the answer more often. This has a major flaw: a correct answer does not guarantee correct reasoning. A model can guess, use a shortcut, or even have flawed logic but still output the right answer. This approach completely fails for tasks like theorem proving, where the process is the product. DeepSeekMath-V2 tackles this with a novel self-verifying reasoning framework: the Generator: One part of the model generates mathematical proofs and solutions. the Verifier: Another part acts as the critic, checking every step of the reasoning for logical rigor and correctness The Loop: If the verifier finds a flaw, it provides feedback, and the generator revises the proof. This creates a co-evolution cycle where both components push each other to become smarter This new approach allows the model to set record-breaking performance. As you can see from the charts above, it scores second-place on ProofBench-Advanced, just behind Gemini. But Gemini isn’t open-source, Deepseekmath-V2 is. The model weights are available on Huggingface under an Apache 2.0 license: https://huggingface.co/deepseek-ai/DeepSeek-Math-V2. This means researchers, developers, and enthusiasts around the world can download, study, and build upon this model right now. They can fine-tune or change the model to fit their needs and research, which promises a lot of exciting math discoveries happening soon - I predict (on no basis mind you) that this will help solve computing problems to start with, either practical or theoretical. Beyond just the math, the self-verification mechanism is a crucial step towards building AI systems whose reasoning we can trust, which is vital for applications such as scientific research, formal verification, and safety-critical systems. It also proves that ‘verification-driven’ training is a viable and powerful alternative to the ‘answer-driven’ method used to this day.