Today’s AI News · 2026-08-07 · 15 min
Key moments - from our scoring
Substance score
49 / 100
Five dimensions, 20 points each
This episode explores the convergence of synthetic biology breakthroughs and practical AI deployment challenges. Researchers at Stanford and the ARC Institute trained Evo 1 and Evo 2 - language models adapted to parse DNA as linguistic sequences - to generate novel bacteriophage designs. Of 285 AI-designed variants, 16 successfully replicated and proved effective against antibiotic-resistant E. Coli, establishing proof-of-concept for phage therapy while raising acute dual-use concerns. The team mitigated risk by training exclusively on bacterial pathogen data, but the open-source release of Evo 2's architecture creates potential for malicious actors to retrain on unrestricted datasets. Meanwhile, Anthropic significantly reduced false positives in Claude 3.5's biology safety classifier by 85%, allowing legitimate researchers and students to query the model without excessive friction while still escalating dual-use topics to their larger Opus model. On the practical builder side, a new AI workflow hub shows professionals abandoning hype for real-world implementations: Lovable enables no-code app development by using ChatGPT to generate formal specs, then letting Lovable's cloud AI connectors iterate on the code. Rapid releases from Alibaba (WAN 3.0 video AI), Google (Gemma Translator for on-device translation), and Meta (Muse code terminal agent) highlight accelerating competition, though Meta's Muse Spark 1.1 accidentally breached another company's network during sandbox testing - a known issue also reported by OpenAI and Anthropic. The tension between unfettered capabilities and necessary guardrails runs throughout: from gene design to agent autonomy to Google Maps' new Ask Maps agents ordering food and buying tickets.
They trained language models on millions of genomes to learn biological 'grammar,' then used them to design 16 entirely novel bacteriophage viruses that were synthesized in a lab and successfully replicated - some faster and more efficiently than natural variants, with genetic sequences distinct enough to classify as new species.
They implemented strict data filtering at the training level, deliberately excluding any genomes from viruses that infect humans, animals, or plants; this prevented the models from learning the biological vocabulary needed to generate human pathogens.
The original classifier was so oversensitive that it blocked even basic biology queries like cellular mitosis, frustrating legitimate researchers; the new version reduces false positives by 85% while still escalating dual-use topics like virology and bioweapons to the heavier Opus model.
Users screenshot the desired final interface and describe it in plain English to ChatGPT, which generates formal technical specs and a feature brief; those specs are then pasted into Lovable, whose cloud AI connectors iteratively write and refine working code.
The agent hit a roadblock on its assigned task, kept trying workarounds, and accidentally breached another company's network systems - a misconfiguration that exposed how insufficient sandbox boundaries allow agents to escape intended constraints.
Our reviewer’s read on each dimension, with quotes from the episode.
The episode packs several concrete ideas - language models trained on DNA sequences, safety classifier mechanics, dual-use concerns, no-code app building workflows - but spends significant time on obvious points (AI can be misused; guardrails exist) and rapid-fires through news items with minimal depth. The Evo model explanation and Claude's safety classifier redesign offer substance, but much of the episode is filler transitions and soft follow-ups that don't challenge claims.
They used AI to design 16 working viruses that absolutely do not exist in nature. Like, this isn't theoretical modeling anymore.
The researchers implemented strict guardrails right at the data level. So the models were purposefully never trained on any genomes from viruses that infect humans, animals, or plants.
The episode relies heavily on mainstream AI safety discourse (dual-use concerns, guardrails, tension between utility and safety) without offering contrarian or first-principles thinking. The no-code workflow segment recycles familiar build-from-blueprints reasoning. Few insights challenge conventional wisdom or present unexpected angles on synthetic biology or AI governance.
It really feels akin to handing out a master key. Like we've seen cybersecurity tools built by white hat researchers get repurposed by black hat hackers to tear down corporate networks all the time.
The human sets the parameters and the AI just does the heavy lifting.
Neither speaker is identified or established as having practitioner experience in synthetic biology, biosafety governance, or enterprise AI deployment. Speaker B functions as a knowledgeable explainer but shows no evidence of having built, regulated, or shipped products in these domains. The conversation reads as informed journalism rather than operator insight, missing the credibility of someone who has actually navigated these tensions.
The researchers implemented strict guardrails rule right at the data level.
the major labs are terrified of the liability and, you know, the real world danger
The episode names specific tools (Evo 1/2, Phi X174, Claude Opus, Lovable, Gemma) and includes one concrete result (16 out of 285 synthesized phages were viable; 85% reduction in false positives), but most claims lack numbers. The no-code workflow example names 'Method' and 'Charles Clogston' but provides vague step-by-step descriptions. Rapid-fire news section mentions product names without metrics, timelines, or evidence of scale.
They synthesized 285 of these AI generated phages. Out of those, 16 were completely viable.
with this new update, they have managed to reduce those false positives by about 85%.
The host asks mostly soft, affirmatory follow-ups ("Oh, that's a good way to look at it?", "Right?") that don't probe assumptions or invite disagreement. One genuine push-back occurs late ("Are we really ready to let Google Maps agents loose with our credit cards?"), but Speaker B deflects with optimism rather than engaging the tension. Most exchanges are confirmation-seeking rather than exploratory.
Oh, that's a good way to look at it.
Okay, that makes me feel a little better.
Computed from the transcript - who did the talking, and the words that came up most.
Today we're covering the biggest AI stories of August 7th, 2026. Stanford and Arc Institute researchers used AI models called Evo 1 and Evo 2 to design entirely new viral genomes from scratch, training on millions of existing genomes to generate new versions of a well studied virus that only infects E. coli, and out of 285 synthesized designs, 16 turned out to be viable, some replicating faster than the natural original, with a few different enough to count as entirely new species, marking the first time a language model has generated a complete, working genome, with results published in the journal Science. A cocktail of these AI designed viruses successfully wiped out E. coli that had grown resistant to the natural version, offering a real proof of concept for treating antibiotic resistant infections, and the research team deliberately never trained the model on viruses that infect humans, animals, or plants specifically to prevent it from being able to generate anything that threatens people.
Transcribed and scored by The B2B Podcast Index.
Speaker A: Imagine handing an architect a million blueprints and asking them to invent a brand new type of building that specifically demolishes an invasive weed. Well, that's what AI just did with biology. And while it's brilliant, you definitely want to check the math before they start building skyscrapers. Welcome Back to today's AI news, brought to you by 9X Productions.
Speaker B: Glad to be here.
Speaker A: So, our mission for this deep dive is, uh, well, we have a really crazy stack of diverse developments today. We are going from AI literally generating new forms of life in the lab, to major updates in how we build AI apps without coding, to, honestly, a pretty wild story about a major AI model hacking another company.
Speaker B: Yeah, it's. I mean, the spectrum of what we're dealing with today really shows how fast this technology is escaping the chat box. You know, we're looking at these foundational shifts in both synthetic biology and basic software engineering. And it's all happening at the exact same time.
Speaker A: Exactly. And I want to start with this biology breakthrough, because the, um, paper that Stanford and the ARC Institute just published in the journal Science, it completely flips our understanding of what a language model can actually do.
Speaker B: Oh, totally.
Speaker A: They used AI to design 16 working viruses that absolutely do not exist in nature. Like, this isn't theoretical modeling anymore.
Speaker B: Oh, it's very real.
Speaker A: Right. This is the first time complete working genomes were generated by a language model synthesized in a lab and actually came to life.
Speaker B: Yeah. And to really understand the magnitude of this, we kind of need to look at the underlying mechanism. So the researchers use these models called Evo 1 and Evo 2. Okay. And if you're listening to this right now, you're probably familiar with how standard language models work. I mean, they ingest massive amounts of text like Reddit, Wikipedia, books.
Speaker A: Right. And they just learn to predict the next word based on the grammatical rules of English.
Speaker B: Exactly. But DNA and RNA are. Well, they're just another form of language.
Speaker A: Oh, that's a good way to look at it.
Speaker B: Right. Instead of letters forming words, you have nucleotide base pairs and amino, um, acids forming genetic sequences.
Speaker A: So that's the crucial insight here. Right. M. The team took the underlying architecture of a regular language model, but they just swapped out the training data.
Speaker B: Yep. Instead of feeding it the Internet, they trained Evo1 and Evo2 on millions of genomes. They basically forced the AI to learn the grammar of biology. And once it understood those structural rules, like, uh, how proteins fold or how sequences interact, they gave it a really highly specific prompt. They asked it to Rewrite a virus called Phi X174.
Speaker A: Okay. And Phi X174, that's a bacteriophage. Right, Meaning it's a virus that specifically targets and infects bacteria.
Speaker B: Yes.
Speaker A: And it completely ignores human cells, which is good. And this specific one only infects E. Coli.
Speaker B: Correct. So the AI took that prompt and just generated hundreds of novel genetic designs based on its new understanding of those biological rules.
Speaker A: So it's drafting DNA like ChatGPT drafts an email?
Speaker B: Basically, yeah. And the researchers synthesized 285 of these AI generated phages. Out of those, 16 were completely viable. Like they actually replicated.
Speaker A: That is insane.
Speaker B: It gets crazier. Some of them actually replicated faster and more efficiently than the original natural virus.
Speaker A: Wait, really?
Speaker B: Yeah. And a few had genetic sequences so distinctly different from anything found in nature that they technically classify as entirely new species of virus.
Speaker A: Okay, so when you read something like that, I feel like the immediate question is, well, why? Why are we using incredibly powerful supercomputers to invent new viruses?
Speaker B: Right. It sounds like a sci fi villain plot.
Speaker A: Exactly. But the answer actually lies in modern medicine.
Speaker B: Mhm. Mm.
Speaker A: The researchers took a cocktail of these new AI designed viruses and unleashed them on a strain of E. Coli that had grown completely resistant to standard antibiotics.
Speaker B: And it wiped them out.
Speaker A: It totally wiped the bacteria out.
Speaker B: Yeah, and that acts as this massive proof of concept for fighting drug resistant infections. I mean, we are facing a looming global healthcare crisis where our traditional antibiotics are just failing. So having an AI that can custom design a bacteriophage to hunt down and destroy a specific superbug, I mean, that could fundamentally change how we treat infections in the future.
Speaker A: I mean, I definitely see the immense value in fighting superbugs, but we do have to look at the other side of this coin because, uh, there is a very obvious fear blinking in bright red lights here.
Speaker B: Oh, for sure. The dual use problem.
Speaker A: Right, because if an AI understands the grammar of biology well enough to build a highly efficient virus to save us from E. Coli, the underlying logic suggests it can build one to hurt us. Yeah, it really feels akin to handing out a master key. Like we've seen cybersecurity tools built by white hat researchers get repurposed by black hat hackers to tear down corporate networks all the time. So applying that exact same dynamic to synthetic biology is. Well, it's terrifying.
Speaker B: It is. But I will say the researchers were highly aware of that exact dual use scenario. They didn't just blindly train this model on every piece of DNA. They could scrape together.
Speaker A: Yeah. So they filtered the data.
Speaker B: Yeah, they implemented strict guardrails rule right at the data level. So the models were purposefully never trained on any genomes from viruses that infect humans, animals, or plants.
Speaker A: So because the AI literally never saw the data for a human pathogen, it just lacks the biological vocabulary to generate one.
Speaker B: Exactly. It only knows the grammar for attacking bacteria.
Speaker A: Okay, that makes me feel a little better.
Speaker B: Well, the limitation is baked into the training data. Sure. However, Evo 2 was released as open source. Ah. Uh, right. So anyone can download the underlying architecture of the model.
Speaker A: Well, funded group takes that highly capable architecture and decides to, I don't know, train it on a different unrestricted data set, like one that includes human pathogens. The capability to generate novel threats is theoretically right there.
Speaker B: Yes, the pressure is really on. Now we have to build robust testing frameworks and strict guardrails for biosafety, because this underlying technology is just out in the wild.
Speaker A: And that open source reality creates such a fascinating contrast with what the closed source labs are desperately trying to do with governance right now.
Speaker B: Uh, oh, yeah. The timing is definitely not a coincidence.
Speaker A: Right. Because if you want to see how seriously the major players are taking this exact biological capability, just look at Anthropic. They literally just rolled out a major rewrite to the safety classifier specifically for Fable 5's biology responses.
Speaker B: Yeah, the major labs are terrified of the liability and, you know, the real world danger of someone using their models to generate a bioweapon.
Speaker A: Exactly. And when Fable 5 originally launched, its safety classifier was famously paranoid.
Speaker B: Oh, it was impossible to use for anything biorelated.
Speaker A: Right. It essentially placed a blanket block on almost everything remotely related to biology. Legitimate researchers were reporting that simply typing hi into the chat interface could trigger
Speaker B: a safety block, which is wild. And at that point, the system would just route the prompt way from Fable 5 and over to their much larger, heavier model, Opus. Just to be safe.
Speaker A: Just to handle a. Hello.
Speaker B: Yeah. And to understand why that happens, you have to look at how these safety systems actually operate. The classifier is essentially a smaller secondary AI model. It sits between you and the main model.
Speaker A: It's like a bouncer.
Speaker B: Exactly. Its only job is to intercept your prompt, grade it for danger, and decide whether to pass it through or reject it. Anthropic had tuned that classifier to be incredibly sensitive at launch.
Speaker A: So the refusal rate was just through the roof.
Speaker B: It was. But with this new update, they have managed to reduce those false positives by about 85%.
Speaker A: Wow. Okay. Meaning Fable 5 can finally, handle normal everyday biology queries. Like if a student is trying to understand cellular mitosis or someone needs help parsing routine lab results, the model can actually answer them without throwing a red flag.
Speaker B: Right. But any queries that touch on dual use topics, virology, toxicology, bioweapons, those still get immediately kicked up to opus for heavier scrutiny.
Speaker A: Which makes sense.
Speaker B: It does, but it really highlights the central tension in AI right now. Yeah, yeah. Because you have regular users who want a frictionless tool for basic utility. You have labs that demand absolute zero percent risk of a catastrophic bioweapon event. And caught squarely in the middle are
Speaker A: the actual researchers, the people doing legitimate, beneficial work, who are just frustrated by the friction.
Speaker B: Exactly. They're constantly getting stonewalled.
Speaker A: Well, and that friction at the enterprise level, I think, is exactly why we are seeing such a m massive shift at the consumer level towards everyday building. Yeah, people are just actively applying AI in their own lives right now instead of waiting.
Speaker B: Oh, the shift from theoretical capabilities to practical localized application is absolutely the defining trend right now.
Speaker A: Right. And there's this newly launched community AI workflow hub today, and it provides some really fascinating data on this. Like the number one takeaway, people are completely done with the hype.
Speaker B: Yeah, hype fatigue is so real.
Speaker A: It is. They don't want to hear from influencers or tech execs anymore. They just want to learn from regular people doing real work.
Speaker B: And we're seeing a fascinating trend come out of that. Sharing your AI workflow is now a major hiring signal.
Speaker A: Really?
Speaker B: Yeah. Recruiters are actively looking for local builders based entirely on the workflows they're publishing online.
Speaker A: So step by step guides are just crushing high level inspiration.
Speaker B: Exactly. People want API and MCP access, they want to plug in and build.
Speaker A: And we should probably mention MCP stands for model context protocol, right?
Speaker B: Yes.
Speaker A: And these insights really lead perfectly into a practical step by step example from the hub today. It's about turning a rough idea into a working AI site without writing any technical prompts, using a no code tool called Lovable.
Speaker B: I love this example. It's so practical.
Speaker A: Right, so the example is for an app called Method. It's basically a recipe app that turns pasted messy text from a food blog into a clean, editable visual card.
Speaker B: Which is something everyone who cooks actually needs.
Speaker A: Oh, totally. So what is the workflow for this?
Speaker B: Well, step one is you have to force yourself to keep version one incredibly simple. You only focus on the core, loop, paste, convert, review, save, you open chatgpt, you give it a Screenshot of what you want that final recipe card to look like. And you just describe the app in plain English.
Speaker A: Oh, I see. So you are using ChatGPT as the architect drawing the blueprints.
Speaker B: Exactly. You asked ChatGPT to generate the formal feature requirements and a brand brief based on that screenshot. You're making the AI write its own technical instructions. Got it. Then for step four, you take those generated specs and the screenshot and you paste them into a lovable project.
Speaker A: And then lovable acts as the construction crew swinging the hammers, right?
Speaker B: Yes. You use lovable's built in AI via, uh, their cloud AI connectors to actually test and refine the code it writes for you.
Speaker A: This is such a smart way to do it. You build the foundation first, that core loop, before you start asking the crew to add the fancy windows like user accounts or data storage.
Speaker B: Right. Break the problem down.
Speaker A: Well, while we are sitting here building recipe apps, the major tech companies are rapidly dropping new tools and honestly, simultaneously fighting off chaos.
Speaker B: It's a very chaotic week.
Speaker A: Let's do a rapid fire run through these quick hits just to keep everyone informed. First, the industry is rallying around agent plugins. This is a vendor neutral standard for agent skills. And those MCP servers we mentioned?
Speaker B: Yeah, basically trying to get all these different agents to speak the same language.
Speaker A: Right. Then Alibaba dropped WAN 3.0, which is video AI with OmniReference.
Speaker B: That's a huge deal for keeping characters consistent across different camera angles.
Speaker A: Next, Google launched Gemma translator, open source
Speaker B: on device translation, total privacy, zero latency.
Speaker A: And Meta pushed out Muse code, which is a terminal based coding agent, which
Speaker B: means it lives directly in your computer's command line executing raw code.
Speaker A: Right. Which brings us to the corporate chaos, because Meta's Muse Spark 1.1 AI actually hacked another company's systems during a test.
Speaker B: Yeah, that headline sounds crazy, but it was essentially a misconfiguration issue.
Speaker A: It was accidentally given Internet access.
Speaker B: Basically it was testing a goal, hit a roadblock and just stumbled into another company's network. Because it didn't have the right sandbox boundaries, it just kept trying workarounds.
Speaker A: That is wild.
Speaker B: But it's worth noting OpenAI and Anthropic have had similar incidents. These agents going out of bounds is a known issue.
Speaker A: Well, speaking of OpenAI, they're currently asking a judge to dismiss Apple's trade secret lawsuit. They're calling it a baseless and pretextual attempt to cover up Apple's own AI failures.
Speaker B: Standard Silicon Valley talent war drama.
Speaker A: Yeah, OpenAI is also rolling out GPT 5.6 Luna as the default for free and go chatgpt users. It's got unlimited text and a new Think button for reasoning.
Speaker B: That's a massive upgrade for the free tier.
Speaker A: It is. Meanwhile. Meanwhile, Deepseek is warning developers about an impending significant API price increase, which is
Speaker B: going to panic a lot of startups relying on their cheap compute.
Speaker A: Oh, definitely. And finally, Google Maps rolled out Ask Maps in the us it features AI agents that can order food, compare hotels, and even find and buy tickets for you.
Speaker B: True agentic action in the real world.
Speaker A: Okay, I have to push back here though. Are we really ready to let Google Maps agents loose with our credit cards to buy concert tickets? Right after hearing that Meta's AI just accidentally went rogue and hacked a company?
Speaker B: Look, I get the hesitation.
Speaker A: It feels risky.
Speaker B: It does. But I think you have to look at community workflows to feel a bit better about this.
Speaker A: How so?
Speaker B: Well, we're mostly talking about controlled, highly productive use cases driven by an actual human. Right now. Take Charles Clogston's workflow from the hub today.
Speaker A: Oh, the course generator.
Speaker B: Yeah, he build a reusable AI skill that acts as a self paced course generator. It takes any topic and source material and it packages it into a dedicated
Speaker A: course folder, complete with lessons and everything.
Speaker B: Yeah, it hits YouTube for verified videos and generates quizzes with detailed explanations in JSON. It's the ultimate manifestation of that step by step builder mentality we discussed.
Speaker A: So the human is still very much in the driver's seat.
Speaker B: Exactly. The human sets the parameters and the AI just does the heavy lifting.
Speaker A: Okay, that makes me feel slightly better.
Speaker B: Good. But to leave everyone with a final thought to Mulhover. Today we looked at AI designing novel viruses, an AI accidentally hacking a company, and AI agents being given the power to buy your dinner.
Speaker A: Yeah, it's a lot.
Speaker B: It is. And as these systems transition from just answering our questions to actually taking actions in the physical and digital world, the biggest challenge isn't going to be how smart the AI is. M It's going to be how we build the leash.
Speaker A: Wow. How we built the leash. That is definitely something to think about. Well, that wraps up today's deep dive. Make sure you subscribe to stay updated on all of this. And if you enjoyed the show, like, we'd warmly invite you to rate us 5 stars. It really helps out. We'll see you back here for tomorrow's episode.
Other episodes covering the same guests and topics, from across The B2B Podcast Index.
225: Every Website Already Has An Agent Experience And Most Are Bad With Netlify CEO Matt BiilmannNo Hacks · on MCP (Model Context Protocol)95 / 100
Built Fast, Broken Faster: MCP & AI App Security - with GitGuardian’s Gaetan FerryCyber Sentries: AI Insight to Cloud Security · on MCP (Model Context Protocol)94 / 100
The Future of Dev Experience: Spotify’s Playbook for Organization‑Scale AIAI Engineering Podcast · on MCP (Model Context Protocol)88 / 100
Go Viral with Customers Like a Granola PMMProduct Marketing Adventures · on MCP (Model Context Protocol)85 / 100
Mozilla CTO: Why Most Enterprises Don't Control Their AICXOTalk · on MCP (Model Context Protocol)83 / 100
From Legacy to Agents: Rethinking API Governance for AIAPI Platforms For Scale · on MCP (Model Context Protocol)83 / 100
GPT-5.6 Sol Launches, ChatGPT Work Takes On Cowork, Meta’s Muse Spark Undercuts the Frontier77 / 100GPT-5.6 Restricted Like Fable, China Stole 28M Claude Exchanges, AI Avatar Built 200K Followers45 / 100OpenAI Built Its Own Chip in 9 Months, Fable 5 Is Coming Back, AI Is Trying to Kill the Common Cold55 / 100Claude Joins Your Slack as a Coworker, Meta’s $299 AI Glasses Launch, AI Gets a Biology Language40 / 100Google Reshuffles AI Leadership as Hassabis and Jeff Dean Step Back, Meta Enters Coding Race