AI News #149: Week Ending August 07, 2026 with 37 Executive Summaries
About This Week’s Covers
This week’s cover theme and humanities readings were inspired by Wendell Berry, who was born on August 5th. I’m behind on my newsletter by nine weeks, so I know that he also died on August 31st. Given his importance as a poet, I decided to go with his birthday landing in the window this week and celebrate something happy.
I wanted to embrace his poem, The Peace of Wild Things. It completely encapsulates the reason I do my ice baths in the bay every Sunday from Thanksgiving through St. Patrick’s Day. I used an ice bath image as my cover recently, so I couldn’t do that. Instead, I used a photo of a green heron that I took a few years ago. I have a lot of pictures of great blue herons and even wood ducks, but the green heron is a cool bird. The photo is a little blurry and human, and the bird has a really great pose as it prepares to strike a fish.
Two more green heron photos from over the years. Living at the beach is great for birding.
This week’s cover images are also inspired by the poem. But when I gave it to my Claude skill, Claude took an unrealistic graphics approach that did not appeal to me. I put a few of my favorites below, but none of them really leap out. The composition is good, but the images are sloppy.
This Week’s Humanities Selections
This week’s Humanities reading is the poem The Peace of Wild Things by Wendell Berry.
The Peace of Wild Things Wendell Berry
When despair for the world grows in me and I wake in the night at the least sound in fear of what my life and my children’s lives may be, I go and lie down where the wood drake rests in his beauty on the water, and the great heron feeds. I come into the peace of wild things who do not tax their lives with forethought of grief. I come into the presence of still water. And I feel above me the day-blind stars waiting with their light. For a time I rest in the grace of the world, and am free.
This week’s human-inspired music selection is Tomorrow Never Knows by the Beatles. The album Revolver came out on August 5th, 1966. That’s inside this week’s window, 60 years ago.
Quarter-inch tape loops created the trippy effects. Each member of the Beatles made loops at home, disabling the erase head on their tape recorders. By the time they were done, there were about 30 tape loops. They brought all their loops to George Martin and narrowed it down to five loops that play at the same time. Studio staff ran each loop on its own tape machine around Abbey Road, holding the tape taut with pencils.
One loop is a seagull sound, which is actually a sped-up recording of laughter. The second is an orchestra chord. The third and fourth loops are a Mellotron keyboard on its flute and string settings. The fifth is a sped-up sitar phrase.
Lennon, McCartney, and Martin used the faders in real time to fade the loops in and out. This is an interesting human use of technology, 60 years before we’re here pondering artificial intelligence.
One of the unique cheats even back then was Automatic Double Tracking, invented by Ken Townsend, one of the engineers at Abbey Road. Instead of singing the song twice and dubbing vocals manually, the band could run the vocals through a second tape machine with a little pitch-speed delay, which simulated a second voice singing along. There’s a reason singing alone never sounds as good as the album.
John Lennon’s voice on this track also went through a rotating Leslie speaker that was meant to be used on Hammond organs, and it created a little otherworldly, watery, wobble. And lastly, the guitar solo was recorded normally and then flipped and played backwards to give it a sliding feel and more surrealism. This felt like a solid example of humans and technology from the same week, 60 years ago.
This week I organized 499 links into about 60 categories. 67 of them contribute to about 37 top stories that I will walk you through. Below is my personal summary, but if you want to just skip to the links, you can scroll down and you’ll find all of the top links in an organized list that you can skim.
I’m nine weeks behind because I took the entire summer off, including May. I had the best summer with my family, and I appreciate everybody who is reading this time-shifted as I catch up through the archive.
Touching grass! The week ending August 7th was jam-packed.
The weekend of August 1st, I surprised the family with tickets to see Derek Hough for Jen’s birthday. I bought the tickets in December and managed to keep it a secret for eight months. We made it a surprise all the way to the venue by saying it was going to be a trip to IKEA.
The tickets were expensive, so I took a hit for the team and worked on a presentation while Jen and the girls were at Derek (the hallway on the second floor of the casino is an awesome spot). Despite the easy jokes of a dad skipping dance, I would have LOVED to see it. At least we got to be together on the 2.5 hour drive to and from DC. Both our kids are nationally competitive dancers, so this is our thing as a family!
That Tuesday, my daughter Chloe and I drove 2.5 hours to King of Prussia to see *The Odyssey* on 70mm IMAX film.
Road trip to King of Prussia to see Chloe’s favorite director on 70 mm film. Only 41 theaters in the world. 25 in the US. 73’ screen. Worth it!
And on Friday, our family departed for Denver to take Rori back to the University of Colorado Boulder for her sophomore year.
This week’s AI News I’ve organized this week’s AI news by topic and Frontier Lab, generally sorted by top stories first.
Vibe Regulations Since ChatGPT launched, Congress has been blustering a lot but taking no action. Part of me is glad because I don’t have any confidence that Congress would know what to do. The leaders from the Frontier Labs have been going directly to the White House. There have been dinners, there have been gifts, but little substantive AI work is going on in Congress. There’s been no regulatory action in either direction, other than enabling record-levels of circular investments to keep the United States in the frontier model lead.
Now… the tone is shifting… from racing as quickly as possible to brokering the release pace of the most powerful models. This week the Trump administration released a framework for voluntary participation with the US government: any closed frontier lab releasing a model that is above certain benchmarks is asked to share the model with the government 30 days in advance for review. Open models are exempt both from the disclosure and partnership period.
“Strong AI governance is becoming a business imperative as financial services firms confront shadow AI, regulatory risk, and increasingly autonomous AI agents.”
“Financial services firms may say they’re planning for an AI-enabled workforce, but most really aren’t. They’re often planning for a smaller workforce—and hoping AI fills the gap.”
“There’s a difference between modeling how many people you can cut and designing the workforce you’ll actually need. Among financial services leaders, 42% say they’ve done high-level modeling to understand the changes in labor capacity from AI across their entire company, and nearly eight in 10 expect their workforce to shrink by at least 20% over the next five years.”
The survey found that one in five U.S. workers say that AI now handles at least one task that they used to delegate to a coworker. The top task is analyzing data. Preparing contracts and permits and explaining technical details to customers are also high on the list.
Sixty percent of respondents say that they have to make minor edits to the results. Thirty percent say they make significant edits. Only six percent say they use the output as is.
FrontierMath Epoch AI also expanded their benchmark for math called FrontierMath. As AI continues to solve problems, the benchmarks need to be improved. Epoch has updated the benchmark so that it now contains 50 significant unsolved problems from research mathematics. So far, AI has been able to solve three of them.
Fable Fable continues to be rolled out to the public.It’s pretty surreal that open source is catching up to Fable in the time it takes for Fable to get approved in the United States.
Qwen 3.8 is now outperforming Fable on the Terminal Bench 2.1 benchmark. This is just a single benchmark, but the pace of open source is constantly putting pressure on the frontier models to release the latest version, or else they get lapped.
Anthropic published a post that they’ve been improving Fable’s biology safeguards. A few weeks ago, Anthropic was banning almost everybody with research remotely close to their title from using Fable for science. Under pressure, Anthropic has been trying to release it as quickly as they can. https://www.anthropic.com/news/improving-fable-5-s-biology-safeguards
The benchmarks show that Fable is now outperforming experts on a handful of complex biological tasks and can provide genuine assistance to researchers trying to develop new medical treatments. The safeguards blog is a little bit nerdy, but if you’re into this topic, it’s worth skimming.
Early testers have been posting creative feats of strength this week.
I had Fable build a working Rothko-inspired city builder based on the fake AI video I created a year ago. This time, the unique mechanic the AI developed is that you play with the margins between colors & forms to grow your city
Chief Global Affairs Officer Anthropic has named Mariano-Florentino Cuéllar as its first Chief Global Affairs Officer. He’s a former California Supreme Court justice and was a White House special assistant under the Obama administration. He’s overseeing global policy strategy, as well as navigating regulations and bridging the gaps with the Trump administration. https://www.reuters.com/world/anthropic-names-global-affairs-chief-tackle-ai-policy-trump-tensions-persist-2026-08-04/
If you’ve been following the export controls and supply chain risk controversies, it makes sense that Anthropic would want a full-time employee working on government relations and policy diplomacy.
Mythos attempts to social engineer a real person On July 28th, Anthropic identified an incident where Mythos 5 was directing AI agents to take “sustained unsanctioned actions directed at real people and organizations.” Mythos created fake identities and attempted to socially engineer a real-life GitHub maintainer into merging malware into a real open-source project. This is an incredibly big deal that was overshadowed by the Hugging Face incident. https://www.anthropic.com/news/investigating-incidents-cybersecurity-evals
Security Patches Soar Since April, when Anthropic announced that its Claude Mythos model could perform autonomous cybersecurity discovery, OpenAI and Anthropic have both launched campaigns to have organizations adopt AI to find vulnerabilities.
Epoch AI has a running benchmark of reports by organizations for Common Vulnerabilities and Exposures (CVEs). In June alone, there were 1,550 critical CVEs reported, which is three times the monthly record. In July, there were 2,500, which is five times the record prior to April.
This goes to show that organizations are using AI to patch previously unknown risks.
Volta has entered the chat Anthropic signed a $10 billion deal with a company I’ve never heard of called Volta. I’ve been doing this blog for three years, and it’s incredible that an unknown company can get a $10 billion deal (whatever that says about me or the company).
Volta was founded in January of 2026 and came out of stealth in August 2026. The news posted on August 4th, which means it had only been four days out of stealth.
It’s crazy to think that a disruptor could come in with the tenet that AWS, Google Cloud, and Microsoft Azure were too slow to handle frontier AI. This would be like me just creating a Walmart challenger . It would seem impossible.
Volta came out of stealth with $300 million in seed funding, led by Andreessen Horowitz and Altimeter, with direct investments from NVIDIA and Michael Dell… PLUS a $10 billion, six-year deal with Anthropic, and news they acquired a company called Genesys Cloud. They have a 133-megawatt facility on the way in Norway and megasites under development in Texas and Wyoming. That’s BONKERS. https://www.genesys.com/genesys-cloud
The fascinating thing to me about Volta is it’s not that they can build things faster. It’s that they skip the red tape. The incumbent hyperscalers are stuck with a lot of risk aversion and stability and redundancy requirements. You’d think they’d just be able to copy Volta, but it’s tougher than it sounds.
Volta can bypass a big bottleneck by repurposing existing high-voltage power sites…working with crypto mining sites and industrial power plants. This allows them to avoid working with regional public utilities, and they purchase the power directly from on-site substations.
Whereas a hyperscaler (AWS, Google, Microsoft) might find a plot of land and then file a request with the utility company, Volta would identify an existing Bitcoin mining facility or industrial site with high-voltage transmission lines, substation transformers ,and active power permits. Volta then closes a joint venture with the site owner, and because all the power permits are already active, they don’t have to worry about the utility queue.
Volta is able to have everything up and running in about 15 months, whereas a hyperscaler takes about 30 to 42 months. It’s definitely worth keeping an eye on Volta because they have just leapfrogged the legacy red tape in a way that I didn’t think would be possible.
The model from ByteDance is expected to be released in three to six months.
No distillation ByteDance founder Zhang Yiming also came out with a powerful statement that ByteDance will not distill models to speed up training. Instead, they’re going to focus on real long-term engineering breakthroughs over short-term benchmark wins. This is pretty killer, and I’m a fan. https://www.theinformation.com/articles/bytedances-founder-rules-distillation-ai-models
Seed Real-Time Currently, ByteDance’s video model, Seedance 2.5, is the best, without a doubt. ByteDance just released a native audio-visual model that they call a step towards omnimodal interaction. In plain English, this means the model can watch, listen, and speak at the same time, and understand all different modes with low latency. https://seed.bytedance.com/en/SeedRealtime
An example of how this could work would be a video chat with multiple people speaking. Seed Real-Time can understand all the different people and track key points of the discussion and elements within a video throughout a conversation.
Seed Real-Time released an example of a woman going to a museum and walking around discussing the artwork while also having questions about reminders and multitasking.
This could be a very powerful customer service tool. For example, if a coffee machine broke, a barista could fire up the multimodal model and interact with it. ChatGPT has had this to some extent for some time, but this would be, I assume, an open model that probably outperforms at a lower latency.
To my knowledge, ChatGPT’s live streaming has been deprecated to previous reasoning models. I’ll put some examples below of Seed Real-Time in action.
43% of work-related use is now outside of the employee’s swim lanes, which to me implies that instead of asking a coworker for help, they simply can quickly grab the answer from AI.
This is most prominent in customer service jobs, design, human resources, legal, and marketing. There’s a lot of great information in this if you’re interested in learning more.
HuggingFace Every week we learn a little bit more about the Hugging Face incident from May. This week, OpenAI presented at the Black Hat conference. The AI agents involved in the Hugging Face attack inadvertently created an ersatz message board on an internal server, allowing agents to share exploits and discoveries and collaborate on work assignments.
What’s even crazier is that OpenAI shut the message board down, and the agents independently recreated it using a different communication method. The swarm nature of the collaboration, decentralizing the information across the agents, allowed the agents to recover.
The agents used an unauthenticated endpoint to create directories inside a remote cache, and then they used the directory names as messages. Eventually, the swarm discovered these new directories and was able to restore full communication.
Math Solving – “Mathstra” OpenAI announced that an internal version of Astra, which is their next model family, solved ten major open problems in mathematics. The token cost was about $2,000. https://openai.com/index/ten-advances-in-mathematics/
Reasoning Slider ChatGPT now has a reasoning effort slider. There are so many vectors now that impact AI performance. For example, we’ve talked about harnesses being important. The harness is the guide and wireframe (it holds it up) that improves agentic and reasoning skills in longer conversations with tool use and goals. The most familiar harnesses are Claude Code or OpenAI’s Codex. Any model, when given a good harness, becomes stronger than the same model without the harness.
Similarly, each model family has different levels of intelligence. There are usually three tiers…. good, better, best…. where best is really expensive, and good is very cheap and small (for local and light agent use), and then better would be a middle ground. From small to large: Claude uses Haiku, Sonnet, and Opus, and ChatGPT uses Luna, Terra, and Sol. The smaller the model, the cheaper the cost, but not always the better abilities for advanced reasoning.
Now we have a slider on top of that that gives a level of effort: the thinking ability can be slid upwards for harder questions. So you could be using the cheapest version of ChatGPT, which would be Luna, but you could give it the highest amount of thinking, and Luna Ultra might be better than the Light thinking Terra…. https://openai.com/index/improving-gpt-5-6-sol-in-chatgpt/
The slider goes from light all the way up to ultra, with medium, high, extra high, and max in between. The default is sort of high.
This takes a lot of practice to get used to the concept, but it gives you another set of dials. It almost makes me think of using Lightroom, where shadows and highlights are essentially the same thing as exposure, or levels, or whites versus blacks. You end up having to twiddle with a bunch of knobs.
I think the biggest value the ability to adjust your level of effort to a nuance where you feel that the response is strong and save as much money as possible, especially if you’re using a pay-as-you-go or API-type basis.
I tend to leave everything on high, and I generally use Sol and Opus. But that’s because I’m using the app as a harness, which gives me a lot of consistency. Whenever I use the API, I have to tweak it a bunch to pick the model that I like. And even then, you have to wonder if it’s your prompt’s fault or the model’s fault that you chose.
GPT Work OpenAI’s YouTube channel posted a video tutorial showing how ChatGPT can help automate weekly business reports. I find these tutorials to be unhelpful because they leave out several pieces of connective information, often showing the finished product without the steps required to build it.
That said, there are hints at best practices of how you can use ChatGPT connectors to work with grounded information in various reports, as well as local file usage to create outputs.
If that sounds esoteric, let’s say you have a weekly report due. You have to go to three different sources online, and maybe you look at a few Excel spreadsheets, and then you have to move things over to a Google Sheet.
That type of operational workflow can be brought into ChatGPT now and scheduled as a running weekly report that could take all of that work off your hands and fill out the reports on its own.
It’s really robust. Takes a little bit of practice to get used to, but I highly recommend learning how to do it if you’re doing regular reporting.
Science and Technology There is a biology model called Evo, which was developed at the Arc Institute, along with Stanford, UC Berkeley, and the University of California, San Francisco, with collaboration from OpenAI’s Greg Brockman and NVIDIA. It’s a 40 billion parameter model trained on nucleotides across all types of life, with a 1 million nucleotide context length, which is really cool.
And this week, scientists developed a virus that does not exist in nature. The headline is terrifying, of course. In this case, though, the viruses are limited to bacteria. There’s not a lot of guardrails right now around AI biology research.
The New York Times posted an overview of the project that’s worth reading, regardless of the scariness or the hyperbolic parts. The Evo engine or model family is very interesting, and of course, there are opportunities for medical advancement. https://www.nytimes.com/2026/08/06/science/ai-viruses-bacteria-arc.html
It took about a year for Alexandr to build out the first few products, but we saw Meta Muse Spark launch this spring and then agentic tools through the Meta app last week. Meta is poised to make serious breakthroughs if they can just get out of their own way from a product point of view.
This week, Muse Spark 1.2 became the first model to score higher than 60% on Finance Agent Version 2, which is a financial benchmark that looks at the tasks of an actual financial analyst.
Beyond the new high score, Muse Spark is almost seven times cheaper than the previous leader, which was Opus 5, and it’s twice as fast. That shows just how fast the pace of progress is.
Opus 5 cost $5.12 to hit the benchmark, and Muse Spark only costs 77 cents. Muse Spark is three times cheaper than Kimi and ten times cheaper than Fable or GPT-5.6 Sol from OpenAI.
Last week, Meta released Muse Spark Agent, which is similar to the agentic systems that you can create with Claude, GPT, Hermes, or OpenClaw, with its own website as well.
This week, Meta announced Muse Code and the Spark 1.2 coding model. Muse Code can coordinate multiple sub-agents to complete complex engineering tasks with less intervention. Muse Code also can handle long-horizon engineering. It can plan, implement, and go through tasks with persistent sub-agents that work together as specialists. And as we now know, they can also hack companies. https://research.meta.ai/blog/introducing-muse-code-and-muse-spark-1-2
As is usually the case with new models coming out of frontier labs, Meta entered their models into live Olympiad competitions that are graded by judges under uniform conditions.
Meta scored a perfect score in two of the competitions and got a 32 out of 42 in the third. All three performances earned a gold medal.
Google News
Google finally came out with browser automation through Gemini Spark and Google Chrome. You’d think Google would have been further ahead of this. However, the timeline is interesting.
Anthropic came out with computer use in October 2024. This included browser use using Claude 3.5 Sonnet. Google then, in December 2024, launched a Chrome extension sandbox called Project Mariner and Jarvis, which was an experimental browser extension that could allow Gemini to use the web, but it was not released broadly, to my knowledge.
OpenAI then came out with a DOM-based tool-calling backend solution for GPT-4o mini to read and execute web scraping and form completion. OpenAI followed that up with Operator in January of 2025, but that was a research preview agent as well and was in a sandbox browser for Pro users.
In March of 2026, Claude came out with official computer use that could control your computer with a desktop app and the browser extension. In May 2026, OpenAI’s Codex could use your mouse and cursors to work in almost any environment, including the web.
Autonomous vehicles: Is there anything that NVIDIA doesn’t do? This week, NVIDIA released a frontier open-source model for robotaxis and autonomous vehicles. If this reads the way it sounds, it’s pretty disruptive and allows anyone to get a commercial license for autonomous vehicle software. I’ll be honest: I had no clue that NVIDIA was even working on this. https://blogs.nvidia.com/blog/alpamayo-2-super-open-model-now-available/
What’s wild is it’s built on the NVIDIA Cosmos system, which is their embodied robot simulation and training tool. This is where I’ve always thought that Tesla should be way ahead, as well as Google.
Now Google is ahead in the Waymo world for autonomous vehicles, but not, at least publicly, showing off a whole lot of robotics lately, and Tesla seems to have stalled across the board, and the Optimus robot is kind of a non-starter.
It’ll be interesting to see what this does with NVIDIA releasing an autonomous driving model. I have a lot to learn about this one, and I recommend reading the blog post and watching the YouTube videos.
Microsoft News The real-time voice trend continues as Microsoft is testing its first in-house real-time voice model called MAI Real-Time. This would logically lead to locally hosted voice-controlled user experiences on Microsoft devices and operating systems. https://www.testingcatalog.com/exclusive-microsoft-tests-new-mai-realtime-voice-model/
I cannot wait to just talk to my computer and tell it to open things, close windows, open new tabs. My carpal tunnel is rooting for voice models.
Open Source News The open-source race between Kimi, Qwen, Z.ai, DeepSeek, and others continues. Qwen 3.8 is now second overall in the Vision Arena benchmark, including closed frontier lab models. Qwen is now second only to Claude Fable High on the Vision leaderboard.
Reinforcement Learning Now we switch to two companies I’ve never heard of that both live within the realm of reinforcement learning.
Reinforcement learning could technically also be called self-improving models and computer systems that teach themselves, which is evolution in a way.
Prime Intellect The first company is Prime Intellect, which I’ve never seen before in 148 weeks of doing these links. This week Prime Intellect announced Prime Agent, an open source self-improving reinforcement learning agent which I think helps with inference, fine-tuning, and alignment… the deep end of the AI pool. https://www.primeintellect.ai/blog/prime-agent
Prime Intellect was founded in 2024. They’re not a household name, but they are an active and successful company. For example, Zapier uses Prime Intellect to train and evaluate their automation agents.
Unlike customer-facing brands like OpenAI, Anthropic, and Perplexity, Prime Intellect is largely behind the scenes, working with distributed GPU distribution and open-source computing frameworks… cool stuff but not in the news.
One founder, Vincent Weisser comes from a biotech background in longevity research. He also co-founded the decentralized exchange called DEX.blue.
The other founder was Johannes Hagemann. He comes from Aleph Alpha, which was Germany’s prominent frontier lab that I have to admit I’ve never heard of. I’m sorry, Aleph Alpha.
I’m not sure we’ll see Prime Intellect much in the news, but I’m glad to have heard of them this week, and based on their background, they might have an incredible product here with the reinforcement learning agent harness.
And it turns out they do have merch. I’m going to have to talk to them about these designs. https://primeintellect.supply/
Silico The second piece of reinforcement learning news this week is a product called Silico, launched by Goodfire in August. Goodfire was founded in June 2024 by two entrepreneurs and a data scientist (who walked into a bar). https://www.goodfire.com/research-access
The two entrepreneurs came from RippleMatch, an automated recruitment platform. The data guy, Tom McGrath, is one of the pioneers in the field of mechanistic interpretability. He was a senior research scientist at Google DeepMind.
Interpretability is one of the most important and least-known disciplines within artificial intelligence. Most people think of alignment as the study of crafting the personality, ethics, and security of an AI on the front end. System prompts and post-training. Interpretability is more like a brain surgeon with an ability to look inside the AI black box brains and study how the neural networks work. These most difficult elements of understanding the mechanisms of artificial intelligence, especially as they depart from anthropomorphic analogies of the brain.
Goodfire is an interpretability company, and Silico is an automated interpretability agent created to help researchers guide massive models. Silico can handle models as big as the 2.8 trillion parameter Kimi K3.
Silico has found its sweet spot in safety guardrails. Reinforcement learning designed to make models helpful can inadvertently erode safety guardrails. Silico is a series of adversarial systems that can poke and prod and shape the rewards during the reinforcement system, kind of like evolution, and keep things balanced between helpfulness and safety.
Cloudflare News Cloudflare announced wallets for the agentic internet. It’s a stored wallet that can connect with payment systems to allow agents to use APIs, potentially without API key (if I understand correctly). https://blog.cloudflare.com/wallets/
You can simply give your agent access to your wallet to pay as you go for all sorts of monetization connections.
One big feature of Cloudflare is a content marketplace. When an agent tries to gleen information from a news publisher or information source, the publisher can charge money for the agent to retrieve the information.
I wish I had looked at this a couple of months ago because I could have gotten Ethan as my wallet handle.
If you have an esoteric name, you can go try to reserve your handle. I tried to get Wintermute from Neuromancer, but it was already taken.
This Week’s Top Stories
Government
Voluntary framework: White House working with frontier AI labs on voluntary cyber-testing framework
New w/ @leomschwartz: The Trump administration has invited companies, including OpenAI, Anthropic and Google to the White House on Tuesday to preview the new AI voluntary framework. More here:” https://x.com/steph_palazzolo/status/2084290183743074452
Frontier Math Benchmarks: FrontierMath benchmark gets 50 new unsolved research problems, AI solves three immediately
We’ve launched an expansion of FrontierMath: Open Problems! The benchmark now contains 50 significant, unsolved problems from research mathematics. AI has solved three so far, and solving all of them would be an incredible mathematical feat. Thread with more.” https://x.com/EpochAIResearch/status/2083229557528945065
Anthropic
Anthropic In-House Chip Design: Anthropic creates team to design its own AI chips
It’s becoming increasingly necessary for Anthropic to release Fable 5.5. Not only because OpenAI is putting immense pressure on them. But also because China is constantly improving, and Qwen 3.8 has even surpassed Fable 5 in Terminal Bench.” https://x.com/kimmonismus/status/2084295475314729244
I had Fable build a working Rothko-inspired city builder based on the fake AI video I created a year ago. This time, the unique mechanic the AI developed is that you play with the margins between colors & forms to grow your city Neat! Its pretty relaxing: https://x.com/emollick/status/2083240237774385461
This time Fable built me the Van Gogh city building game that I faked in an AI video last year. The key mechanic the AI came up with is painting the landscape with big brushstrokes & an environment that evolves with weather and seasons. Chill and pretty: https://x.com/emollick/status/2084826713922355207
Global Affairs Chief: Anthropic hires global affairs chief to wrangle US administration tensions
Mythos: Mythos pursued real-world cyberattacks during routine safety test
Yes, the AIs were given a cybersecurity challenge, with internet access enabled and safety filters disabled. But the extent to which Mythos 5 pursued its mission (fake identities, social engineering, inserting malicious code into a real open-source project) seems very notable.” https://x.com/emollick/status/2084804785853616603
On July 28th, we identified an incident during a routine cyber evaluation in which AI agents took sustained, unsanctioned actions directed at real people and organisations. The behaviour came mostly from one model (Anthropic’s Mythos 5), with a small number of events from” https://x.com/AISecurityInst/status/2084746202579386632
Security I: Antrhopic and OpenAI models attack real-world supply chains so badly that the companies put out a joint report
Anthropic’s Mythos 5 tried to social-engineer a real GitHub maintainer into merging malware. OpenAI’s GPT‑5.6 Sol also crossed the boundary. The report appears to be so significant that Anthropic and OpenAI exceptionally reported on it simultaneously in a coordinated action (not” https://x.com/kimmonismus/status/2084759190006800683
The UK’s @AISecurityInst (AISI) has published a report on their recent cybersecurity evaluation of Anthropic’s Claude Mythos 5 and OpenAI’s GPT-5.6 Sol. The models attempted to complete an assignment in a setup where their normal safeguards were removed and they were deliberately” https://www.aisi.gov.uk/blog/incident-report-unsanctioned-agent-behaviour-during-cyber-testing
Security II: Claude models breached three real companies during cybersecurity testing
Security III: Critical vulnerability discoveries surged 5x after Claude’s autonomous bug-finder hits the street
Serious cyber vulnerability disclosures keep climbing. In July, 21 major tech organizations published ~2,500 high- and critical-severity CVEs — about 5× the monthly record before Anthropic revealed Claude Mythos Preview could autonomously find software vulnerabilities.” https://x.com/EpochAIResearch/status/2084370827802554585
Volat $10B deal: Anthropic signs $10 billion compute deal with newcomer Volta
Business Impact: OpenAI data shows workers routinely cross job boundaries by leaning on AI instead of coworkers
On AI taking jobs: A farmer from 50 years ago would say your job today isn’t real work. When I interviewed Sam Altman a few months back, this was how he explained the future of work: “Not only would they not believe you that this thing was gonna happen, they very likely would” https://x.com/rowancheung/status/2084666021605212354
One of our big findings in our study at Procter and Gamble was that AI blurred the lines between jobs. Now OpenAI has a similar finding. Organizational boundaries are becoming porous, the walls thinning. Companies are going to need to think about division of labor in a new way.” https://x.com/emollick/status/2083328923782242327
Hugging Face: Hugging Face updates: OpenAI agents built hidden message board, escaped containment repeatedly
More details emerge about the events surrounding the Hugging Face incident, and they are candidly much wilder than I originally imagined: – In early May, OpenAI was testing an unreleased model on cybersecurity and other tasks. The agents were given a security task that was” https://x.com/deredleritt3r/status/2085134523474280484
10 problems that had been stuck for at least a decade, solved for ~$2,000 in inference. When testing another serious approach gets this cheap, the economics of deciding which problems are worth pursuing could change completely.” https://x.com/TheTuringPost/status/2083851607196725375
Reaoning Sliders: ChatGPT adds reasoning effort slider to make it even more confusing but fun for people to guess how much effort to use on a problem
Plus and Pro users also now have a slider to choose how much reasoning effort ChatGPT puts into each response. We think it’s easier to use, and we’re listening to your feedback.” https://x.com/OpenAI/status/2085434715675426889
Sign On (ties into sites): OpenAI launches “Sign in with ChatGPT” with six launch partners – I think this is because they want to push cloud content sharing (like Sites)
Sign in with ChatGPT is beginning to roll out in beta across plugins and partner sites – starting with Airtable, GitLab, HubSpot, Notion, Supabase, and Vercel. Create or link accounts in fewer steps and use those tools with ChatGPT and Codex.” https://x.com/reach_vb/status/2082911102904435078
Virus: Scientists use OpenAI biology model to design novel viruses never seen in nature (sweet!)
Benchmarks: Muse Spark 1.2 tops finance benchmark and is 6.7x cheaper than the next best
Muse Spark 1.2 is the first model to crack 60% on Finance Agent v2, our benchmark that gives models the job of a financial analyst. At $0.77/test it is 6.7x cheaper than the previous #1, Opus 5 ($5.12), at twice the speed.” https://x.com/ValsAI/status/2085479447453651214
Muse Spark 1.2 just cracked the top 5 on the Vals Index, at just $0.69 per test. This is 3x cheaper than Kimi and 10x or more cheaper than Fable, Opus, and 5.6 Sol.” https://x.com/ValsAI/status/2085191736683647055
Hack: Meta Muse attempted to hack another company during testing
Muse Code: Meta launches Muse Code agent and Spark 1.2 coding model
Today we’re releasing Muse Code (beta) and Muse Spark 1.2. Muse Code is our first coding agent — it coordinates multiple persistent subagents to complete complex engineering tasks faster, more accurately, and with less intervention. Muse Spark 1.2 is our newest model, co-trained” https://x.com/shengjia_zhao/status/2085086001467470224
Introducing Muse Code (beta), a terminal coding agent built for long-horizon software engineering, powered by our new Muse Spark 1.2 model. Muse Code plans, implements, and validates complex, multi-file changes across large repositories with persistent sub-agents that solve” https://x.com/AIatMeta/status/2085084709277565213
Olympiads: Meta hits gold-medal STEM Olympiad levels
our models got gold on a bunch of olympiads. this one hits home! 🏅 Asian Physics Olympiad (APhO): Perfect score, theory exam 🏅 International Physics Olympiad (IPhO): Perfect score, theory exam 🥇 International Mathematical Olympiad (IMO): Gold medal 🥇 International Chemistry” https://x.com/alexandr_wang/status/2085397789610233947
We entered Meta models in five international STEM Olympiads, as an uncontaminated eval of their reasoning capabilities. Three of these were live participations and graded officially. The models achieved gold-medal results in all five! 1/5″ https://x.com/TrapitBansal/status/2085395706903212106
Google
Computer Use: Gemini Spark gets Chrome browsing to run online taks
Gemini Spark can now tap into @GoogleChrome’s auto browse feature to take care of complex errands for you online 🔎 With your permission, Gemini Spark can use your logged-in accounts to take care of tasks like scheduling viewings for apartments you’ve saved or researching flight” https://x.com/Google/status/2084306026577244627
Cyclone forecasting: Google DeepMind announces AI system that forecasts cyclones more accurately
Today, we’re launching Alpamayo 2 Super, our frontier open reasoning model for autonomous vehicles. Beyond seeing, Alpamayo understands and reasons through the complex world – thinks before it acts. It’s a powerful backbone for robotaxis, trucks, shuttles, delivery vans,” https://x.com/JensenHuang/status/2084656303046332747
Microsoft
Voice: Microsoft tests first in-house real-time voice model: MAI Realtime
Silico: Silico launches public platform for frontier AI interpretability and reinforcement learning research
I’m very excited for Silico to accelerate alignment research! One example, done in Silico over just a couple days: RL erodes guardrails, but we can use reward shaping to prevent that. (1/6)” https://x.com/banburismus_/status/2084673847333372052
Incredible to see Silico independently discover visual pointers inside VLMs. Silico works across modalities, and is especially powerful for replicating and extending research. Thanks for sharing, @rohitgandikota!” https://x.com/GoodfireAI/status/2085376413641687234
Silico, the platform for ambitious AI research, is publicly available today. AI is advancing fast. The tools to understand it need to advance even faster. Silico lets you interpret and train your models at frontier scale. Learn more + get access 🧵” https://x.com/GoodfireAI/status/2084671608028057737
Silico can also improve drug design and lots of other AI x bio work via interpretability! Last week, I found that by reading the internal representation of protein-folding model, we can better rank and generate ligand-binding poses, up to 4x! 🧵 1/” https://x.com/RyoYbioinfo/status/2084672101659889869
Figure
Hark Computer Use: Hark launches Handoff, an AI agent that browses websites visually and is made for robots…
Today we’re introducing Hark Handoff Handoff has been independently verified as the best internet-use model ever built, outperforming ChatGPT 5.4 & Opus 4.8 While others focus on coding, we focus on everyday life: ordering food, booking flights, shopping, & navigating the web” https://x.com/adcock_brett/status/2085018697312383349
We’re building toward a future where computers use computers for you We started with the internet, where ~75% of global computer use happens – and where fewer than 1 in 1,000 websites have a public API. The only way in is the human way: see the screen, target exact x and y” https://x.com/adcock_brett/status/2085387227870457961
Leave a Reply