ai
Fast Company called AI For Humans 'the most entertaining way to learn about artificial intelligence.' Every Thursday, hosts Kevin Pereira and Gavin Purcell break down the biggest AI news, tools, and breakthroughs, so you can stay ahead without losing your mind. Whether you're AI-curious or deep in the game, this is your weekly reset on everything that matters in artificial intelligence.
Google has previewed its second flagship frontier model in two weeks, named Gemini 4 Carbon, currently available to users on the $100-a-month Google AI Ultra plan. According to Gavin Purcell, the model is targeted at enterprise cyber defense, banks, and government contracts to assist with pen testing and security.
Google DeepMind has previewed Genie 4, which the hosts describe as its most realistic and interactive world model yet. Users can use keyboard controls or image prompts to move through real-time generated simulations, and the model can also be used to train physical robots in tasks like pipetting.
OpenAI has upgraded its GPT-6 model lineup to version 6.1, introducing the Terra tier beneath Sol. Kevin Prair notes that Terra is designed for fast computer use tasks, such as generating 3D models and assets in Blender.
Jack Dorsey's company Block has developed a new AI audio model called Dessa, which generates highly realistic conversational speech with distinct accents and human-like emotional delivery. The model was notably trained on a relatively small compute setup compared to massive clusters.
OpenAI published research detailing breakthroughs in math, physics, and computer science, specifically challenging the exponential time hypothesis (ETH). The hosts explain that this suggests certain complex computational limits may be solvable without a proportional increase in time.
Kevin and Gavin discuss the massive wave of cyber attacks targeting a long-standing PHP vulnerability. According to Matthew Prince from Cloudflare, thousands of exploitation attempts surged to tens of thousands within 45 minutes of the announcement, ultimately exceeding a million attempts.
Gavin shares that Anthropic partnered with Citrini Research using Claude compute to design proteins aimed at developing an antidote for scorpion stings. The project successfully generated potential protein building blocks for future treatments.
Gavin highlights the launch of Flux 3 Max and WAN 3 video generation models on FAL. The new WAN model offers capabilities comparable to Seedance 2.5 at roughly one-fifth of the price, significantly reducing generation costs for creators.
Kevin explains a project by Synthetic Beef that scraped over 100,000 Steam game descriptions, mapped them into vector space via embeddings, and identified gaps containing completely unrepresented game concepts.
Recent advancements in AI, particularly with tools like Opus 5.5, are allowing users to decompile and mash up existing video games. This technology enables players to run elements from one game within another, such as placing Minecraft's TNT in Cyberpunk, leading to a surge of creative, albeit legally dubious, game modifications.
The AI model Opus 5.5 is being hailed as a significant step towards Artificial General Intelligence (AGI), particularly after an AI agent named Fig autonomously created a 27-minute documentary. While the AI handled scripting, generation, and editing, the human creator still had to provide executive production notes.
OpenAI's Dev Day event featured the announcement of 'Dots,' personal AI agents designed to perform tasks for individuals. However, some critics note that Dots' functionality is similar to existing AI tools like Muse and Grokbot, and question its unique value proposition, particularly given its limited availability to paid subscribers and organizational challenges within the ChatGPT app.
OpenAI has reportedly delayed the release of Astra 6.1, citing safety concerns and a belief that the model is not yet safe enough for public release. This delay comes as OpenAI's Dev Day was overshadowed by competing AI models, including Opus 5.5, which some believe outperforms Astra 6.1.
An AI agent named Fig, powered by Opus 5.5 and Runway's MCP, autonomously generated a 27-minute documentary about superintelligence. The AI handled scriptwriting, shot selection, editing, and music, though human executive producer notes were still required. This demonstrates a significant leap in AI's capability for complex media creation.
Several new AI video and visual tools have launched, including Ideogram 4.5, which is reportedly strong in editing still graphics and has seemingly 'solved' AI photo editing. Kling 4.0 offers advanced video generation capabilities, including 32-second 10-bit HDR videos and the ability to create storyboards. ElevenLabs v4 enhances voice generation with more natural expressions and emotions.
Prominent AI figures, including Jensen Huang, Dario Amodei, and Mark Zuckerberg, met with the President to discuss the future of AI and superintelligence. Jensen Huang's prior comments about AI inflicting 'enormous amount of pain and suffering' for long-term benefit have drawn criticism, with concerns that those in power will not experience this 'pain' themselves.
A 'Succession'-inspired parody series, titled 'You Are Not Serious People,' was created using Opus 5.5 and various AI tools. The series, featuring AI-generated characters of industry leaders like Jensen Huang as Logan Roy, demonstrates the increasing sophistication of AI in generating complex video content with multiple shots and nuanced performances.
Anthropic has released its Opus 5.5 AI model, which hosts are calling "probably the best AI model we've ever used." It is noted for its strong coding capabilities, impressive graphics generation, and clear, effective writing, outperforming previous versions and rivaling even their flagship models.
OpenAI has launched two new AI models this week, described as both powerful and affordable. The hosts have extensively tested these models and plan to discuss their strengths and weaknesses.
The release of a new Grok model was briefly mentioned alongside new offerings from Anthropic and OpenAI. Details about this new Grok model were not elaborated upon in this segment.
A common observation in the AI space is that models can degrade in performance after their initial release. This phenomenon is attributed to factors like scaling up compute resources and managing costs, leading to 'dialing in' the model based on global usage.
The discussion highlights that beyond raw power, the usability and conversational fluidity of AI models are crucial. Opus 5.5 is specifically praised for its ability to engage in more natural, human-like conversation, making it easier and more enjoyable for users.
One of the hosts demonstrated the use of Meshy 7.1 to create a full-blown 3D model of himself, posed and giant-sized. This was done as a personal project, with a promise of a tutorial on how it was achieved.
Kevin, a user of the AI agent Instinct, discovered it identified three recurring charges on his credit card that he was unaware of. Instinct successfully canceled one of these subscriptions and secured a refund, recouping approximately $500 for the user.
A user detailed how the AI agent Instinct assisted in apartment hunting by crawling websites for listings and using reverse image search to identify addresses when not provided. The AI then analyzed Google Street View imagery for specific criteria like shade and proximity to dog-friendly parks.
While using the Instinct AI agent for apartment hunting, a user experienced the agent exposing their phone number and personal email address to numerous agencies. This resulted in a significant increase in spam and raised privacy concerns.
The AI agent Instinct has seen a meteoric rise in valuation, reportedly growing from a $50 million valuation in April to seeking a $10 billion valuation in its current funding round, aiming to raise $1 billion. This rapid growth highlights its increasing integration into users' lives.
The Instinct AI agent includes a 'trusted person' feature that allows users to connect their agents. This was demonstrated when one user's Instinct agent messaged another user's Instinct agent, leading to a conversation between the two agents before the human users were involved.
Despite being currently free, the Instinct AI agent is performing substantial work on the backend for users. This includes research tasks and creating Google Sheets, indicating a robust capability even without a paid subscription model yet.
OpenAI's latest model, GPT-6 Astra, demonstrates significant advancements in computer use, capable of performing complex tasks like 3D modeling, controlling cameras, and generating music. However, users report that its computational power comes with a very high cost, quickly depleting token limits even on expensive plans.
Users of OpenAI's GPT-6 Astra are frequently encountering token limits, leading to a need for 'resets' to continue using the model. This has become a common topic of discussion, with some users reporting receiving unexpected resets from OpenAI staff, likening them to a 'Santa Claus' figure.
OpenAI's GPT-6 Astra is enabling individuals with no programming background to create intricate 3D environments using Blender. Previously, such tasks required weeks of work by skilled artists, but Astra can now generate these environments, including rigging, texturing, lighting, and camera work, in a fraction of the time.
GPT-6 Astra is revolutionizing computer use by enabling AI to operate software with human-like proficiency, and in some cases, surpassing human capabilities. This includes tasks such as playing games, 3D modeling, controlling cameras, and composing music, demonstrating a significant leap in AI's ability to interact with and utilize computer systems.
OpenAI's GPT-6 Astra is pushing creative boundaries, moving beyond simple benchmarks like generating images of 'Will Smith eating spaghetti.' The model is now capable of complex tasks such as creating 3D animations, like a 'Billy Jean dance' in Blender, and transforming 2D text adventure games into 3D experiences.
A special integration is being developed to allow Astra agents to communicate with Blender more efficiently, significantly reducing content generation times. Benchmarks show generation times cut from 15 minutes down to approximately three minutes.
Mark Zuckerberg, in a recent essay, outlined his vision for a positive AI future, emphasizing the potential of superintelligence. He also announced Meta's commitment to open-source AI, with plans to release weights for models like 'music glimmer' and 'Muse Spark'.
The podcast discusses rising opposition to AI data centers, noting a 12% increase in opposition over four months. This sentiment is described as 'divisive' and is a concern both globally and within the United States.
The hosts observe a 'devil and angel' dichotomy in AI's current state, with significant advancements occurring alongside increasingly negative public perception and incidents. They question whether the public truly wants the trajectory AI is taking.
The AI video generation landscape is seeing significant advancements with the release of Seedance 2.5, MiniMax H3, Juan 3.0, and Flux 3. Tim Simmons notes that while Seedance 2.5 offers 30-second outputs and extensive reference capabilities, MiniMax H3 is notable for its ability to run locally on PC hardware.
Seedance 2.5 marks a significant advancement in AI video generation, now offering up to 30-second outputs and the ability to use up to 50 image references in its Omni model. However, users like the podcast host noted difficulties with generating specific content, such as fashion models, suggesting the model may have limitations or content restrictions.
The extended 30-second output capability of Seedance 2.5 comes at a significant cost, with the host reporting 1,100 credits per generation. However, there is anticipation that pricing will decrease once the API becomes widely available, leading to a more competitive market and potential sales.
AI video YouTuber Tim Simmons created a two-minute neo-noir film titled 'Deathwalks into a Bar' entirely using Seedance 2.5. He noted that while the 30-second outputs are costly, they are useful for dialogue-heavy scenes and that the short film was completed in approximately 15 generations.
The host experienced repeated rejections when trying to generate videos with Seedance 2.5, particularly when attempting to use fashion models and product images. Tim Simmons suggested this could be due to the model's judgment of the input, potential product branding, or its detection of real people or recognizable figures, indicating potential content restrictions.
The upcoming GPT-6 model from OpenAI is rumored to launch in August, potentially being larger, cheaper, and more capable than its predecessors. OpenAI CEO Sam Altman is currently in Washington D.C. discussing AI with the administration.
Anthropic's new AI model, Opus 5, is being described as a capable model that performs well on benchmarks, even surpassing Fable 5 in some areas. However, users report it is very 'token hungry,' leading to rapid consumption of usage limits.
The discussion highlighted a method using 'gauntlet loops' and specific prompts like '/goal' to guide AI models in creating game prototypes. While these tools can generate impressive tech demos and clones of existing games, the focus is shifting to the testing and refinement of game loops, which still requires human creativity.
Ilya Sutskever has reportedly received funding from Nvidia to scale his research, which is currently undisclosed. Rumors suggest this research may involve unlocking continuous learning capabilities for AI models.
A new AI model named Kimi K3 has been released by Moonshot Labs in China, positioning itself as a competitor to leading American AI models. Early comparisons suggest it is comparable to GPT-5.6 Sole and superior to Opus 4.8, though not yet at the level of Fable 5.
A significant claim was made regarding the upcoming release of OpenAI's GPT-5, with one source suggesting it could be available as early as the following week. This news follows the recent reset of Claude Fable 5's subscription date.
The Kimi K3 AI model has shown proficiency in building web-based games, with one example cited as being more creatively interesting than a GPT-5.6 Sole counterpart. This highlights a potential divergence in creative output between different AI models.
The subscription window for Anthropic's Fable 5 model has been extended, with the end date for its availability being pushed back. This development is noted in the context of increased competition in the AI market.
OpenAI has launched its latest suite of AI models: GPT 5.6 Sol, Terra, and Luna. GPT 5.6 Sol is their new flagship model, available to paid users, while Terra and Luna will be accessible for free. The new models are described as fast, capable, and reliable, with Sol Ultra specifically noted for its performance on agent benchmarks.
OpenAI is merging its ChatGPT and Codex applications into a new 'GPT Work' app. This consolidation aims to combine the general capabilities of ChatGPT with the coding and productivity features of Codex, which included integrations with calendars and Google Docs.
OpenAI has released GPT Live, an updated version of their real-time voice model, previously rumored as 'BD' or bidirectional. A key feature highlighted is the ability for the model to interrupt and be interrupted, creating a more natural conversational flow. The model also demonstrated impressive performance in generating exact transcripts of conversations.
Mark Zuckerberg's Meta has launched two new AI developments: Muse and Spark 1.1. Spark 1.1 is described as a new coding agent, with the speaker noting it is 'fast, it is cheap, and it is surprisingly capable.'
Elon Musk's AI venture, xAI, has released Grok 4.5. The model is characterized as being fast, cheap, and surprisingly capable.
There are circulating rumors that OpenAI's next major model, GPT-6, might be released this month. This speculation follows the recent launch of GPT 5.6 Sol and other OpenAI models.
Reactors has released Lingbot World 2, a feature that allows users to live stream their own virtual worlds. This enables creative uses such as embodying characters like a 'walking hamburger on Mars' or within a '90s shopping mall'.
The character George Castanza has been recognized as the AI meme of the week. The details surrounding how this character achieved this status are to be discussed.
The GPT 5.6 Sol Ultra model from OpenAI has reportedly performed exceptionally well on the 'Agent's Last Exam' benchmark. This benchmark is designed to assess what agents will need to do to be successful in the future.
OpenAI's GPT 5.6 models, particularly Sol, are being highlighted for improved design capabilities, demonstrated through five different games showcased in their blog post. These games include a sailing game, a tiny voids game, a museum website, and a clockwork village game.
OpenAI plans to launch GPT 5.6 on a platform called 'Cerebrus' in July, which is expected to significantly increase the model's speed. This move is anticipated to change how users create and interact with AI models due to the improved performance.
Despite improvements in speed and conversational flow, OpenAI's GPT Live voice model exhibited significant flaws. It struggled with basic tasks like counting letters in numbers and failed to adhere to a 'don't tell my wife' instruction, indicating a lack of contextual understanding.
Anthropic has officially released Claude Sonnet 5, a new AI model that promises to deliver performance comparable to their Opus model but at a significantly reduced cost and increased speed. This release follows earlier rumors and anticipation within the AI community.
Fable 5 is slated to return, with its relaunch anticipated for tomorrow, according to discussions on the AI for Humans podcast. The specific details of Fable 5's relaunch, including any changes to sign-in requirements or pricing, remain unknown, but Anthropic has received approval for its release.
OpenAI's new flagship model, GPT 5.6 Soul, has been announced but is currently experiencing limited access, with only 20 companies reportedly having access. The model is facing governmental processes similar to those impacting Fable, raising questions about the future of advanced AI model accessibility.
There's a notable trend of users and companies shifting towards Chinese AI models, such as GLM, due to restricted access to cutting-edge Western models like OpenAI's and Anthropic's. This shift is driven by the perception that Chinese models offer comparable performance at a lower cost, potentially impacting the global AI market share.
The conversation touches upon the increasing trend of 'Know Your Customer' (KYC) requirements for accessing advanced AI models, questioning users about the data they are willing to share. The hosts ponder whether such measures, including potential biometric data like iris scanning, are necessary for security or an infringement on privacy.