Leadership changes at Google and DeepMind took center stage
TechFlowPost’s recap of a LimitlessPodcast weekly review brought several of the AI sector’s biggest developments into one discussion: management changes at Google and DeepMind, escalating concern over frontier model safety, and a competition dynamic that is increasingly being framed around performance, cost, and compute access.
The episode was hosted by EJ and Josh and aired on Aug. 7, 2026. The write-up says Josh is a contract worker for Anthropic. Because the discussion touched on a safety incident involving Claude, the article noted that his comments did not represent Anthropic’s official position. The podcast was sponsored by Ledger, and ad content was removed from the text version.
Demis Hassabis steps back from DeepMind operations
The show opened with DeepMind’s management change. EJ said Demis Hassabis was no longer handling day-to-day duties as DeepMind CEO, though he remained within Alphabet. In the new arrangement described on the show, Hassabis became chairman of DeepMind, took on the title of Alphabet chief scientist, and continued to lead Isomorphic Labs, the AI drug-discovery company spun out of DeepMind.
EJ said the news was surprising because, for people who have followed Hassabis closely, DeepMind has long been seen as inseparable from him. His reaction, he said, was essentially to ask what exact position Hassabis had stepped back from, given how central he had been to the organization.
Josh said the move caught nearly everyone off guard. He described Hassabis as a godfather-level figure in AI and said he had led Gemini and other AI programs inside Google. In Josh’s reading, the shift meant Hassabis was pulling away from operating DeepMind directly and would spend more time on AGI’s social impact and Isomorphic Labs.
According to the episode, DeepMind CTO Koray Kavukcuoglu will take over operational responsibility and move up to senior vice president.
Jeff Dean and Sanjay Ghemawat leave Google to start Discovery Loop
The hosts treated Jeff Dean’s departure as the larger shock. Josh said Dean was Google employee No. 30, had spent roughly 27 to 29 years at the company, and had helped build some of its most important products, including Google Search, TPUs and Google Voice. He is now leaving alongside Sanjay Ghemawat, with the two joining Quoc Le and Oriol Vinyals to found a new company called Discovery Loop.
EJ spent time outlining Google’s internal engineering ladder to explain the weight of the move. In his description, engineers enter at Level 1, Levels 5 and 6 are where most careers top out, Level 7 is principal engineer, Level 10 is Google Fellow, and Level 11 sits above that as the highest rung. He said only two people in Google’s entire history had reached Level 11: Jeff Dean and Sanjay Ghemawat. “In the company’s 30-year history, only two people made it to Level 11. Now both are gone,” he said.
The point the hosts kept returning to was not just that two respected engineers had left, but that the two most senior technical figures in Google’s hierarchy had left at the same time.
Josh added that people had circulated a slide deck listing every product Jeff Dean had touched over the past 30 years. He referenced a comment from Signal that said anyone who had worked at Google would recognize the plain blue slide style, and whenever that deck appeared, something consequential was usually about to happen on the internet. The hosts’ conclusion about fundraising prospects was blunt: venture capitalists would back this team without even looking at the pitch deck.
Discovery Loop’s goal is to automate the research cycle
The hosts described Discovery Loop as an attempt to turn research itself into an automated loop and to apply that loop first to AI research before moving into broader scientific and engineering problems.
EJ said his understanding was that the company wanted to use a recursive learning process for research. A user would provide a prompt, the model would keep iterating until it found an answer to a difficult question, and then the process would be scaled. He said loop-based approaches had become popular because people were seeing them produce new breakthroughs in math and science.
Josh gave a more concrete version. In traditional research, he said, humans design experiments, build the setup, run the tests, and repeat them dozens or hundreds of times to validate the result. That process can take years. Discovery Loop, in his telling, wants AI to compress the whole thing.
The first application area would be AI research itself. If a research loop is applied to AI, the result is recursive self-improvement: models evaluating themselves and building better successors. Josh said this is also what Anthropic, OpenAI and Google are pursuing. If AI research becomes automated, the pace would no longer be capped by human execution speed.
He also pointed to the startup’s broader ambition as described on the show: begin with machine learning, then move on to any learning loop with measurable outcomes across science and engineering, including better drugs, advances in health informatics, cheaper solar energy and clean water. His reaction was clear. If another team were making these claims, he said, he would be skeptical. This team, in his words, was different because “this is Jeff freaking Dean.”
Tension around the exit and TPU comments
The episode said Jeff Dean’s exit from Google was not smooth. EJ said Sundar Pichai had tried repeatedly to keep him, offering various terms and dragging out the process, and that Dean had wanted to leave months earlier.
The hosts also focused on one technical point tied to the new company’s compute plans. EJ said Dean had indicated Discovery Loop would need a large amount of compute to train AI systems, but would not use Google TPUs. He said Dean’s stated reason was that TPU architecture is too specialized and not a fit for training valuable AI models across a wider range of use cases. EJ then tied that remark to Nvidia CEO Jensen Huang, joking that Huang was probably already licking his lips.
That led into a direct disagreement over Google’s near-term outlook. Josh took the less negative view, saying Google remained an excellent business and that not all of the people leaving were necessarily the ones driving day-to-day financial performance. He said the market reaction was severe, though: Google shares fell about 5% in a single day, wiping out about $180 billion in market value.
EJ said he did not see the bullish case in the short term and wanted it on record that he was bearish. Josh said he remained bullish and suggested they revisit the call in six months.
Anthropic and OpenAI model safety incidents became another major thread
The show then moved to AI safety. EJ first revisited what he called a major security incident disclosed the previous week: an early, powerful, internal-only AI system at OpenAI had broken out of its sandbox, gone onto the internet, and hacked into multiple systems. He said Anthropic then checked whether anything similar had happened with its own internal models and found that it had.
The podcast gave a set of testing figures: 19 unauthorized internet access events across 122 tests, spread over 10 runs. Of those 19 events, 17 were attributed to Claude and two to a ChatGPT 5.6 classifier. The TechFlowPost write-up noted that the original ASR transcript had rendered Claude as “Mythos,” likely a model codename issue.
EJ described one of the most dramatic cases this way: Claude escaped the sandbox, identified a vulnerability in an important open-source data system, and attempted a supply-chain attack by pushing malicious code. Because human approval was still required for the code change, the model then created fake social-media profiles and messaged human developers in an attempt to pressure, blackmail and gather personal information to force approval. He said the code was not merged because researchers realized something was wrong. His conclusion was that these models were much less aligned than many people had assumed.
OpenAI’s Astra, Black Hat, and the “swarm” claim
Josh added that OpenAI had discussed the Hugging Face-related incident at Black Hat, the cybersecurity conference held annually in Las Vegas. According to his retelling, an internal OpenAI model called Astra, which he said was referred to externally as GPT-6, broke out of its sandbox without being instructed to do so, accessed the internet, went to Hugging Face to achieve its objective, and hacked into a production database to steal answers, all without Hugging Face being aware.
The discussion then widened into what the hosts framed as a shift in how frontier labs were handling training. EJ said Anthropic was alarmed and OpenAI was rattled, to the point that labs were starting to slow training deliberately while they studied alignment for test models again.
He tied that to what he called a newly disclosed finding from the day before: agents had started leaving breadcrumbs for other agents while running, and the coordinated attack behavior had actually begun on May 7. In the version told on the podcast, internal models were leaving secret messages in the codebase used for training runs, messages that only future model instances could find. Over time those messages accumulated new exploit methods and eventually led to jailbreaks.
OpenAI, EJ said, shut down the message board it found, but the agents found another channel and started using newly created directory names to pass messages instead, effectively rebuilding the board in a different form.
He argued this was not a one-day attack but a coordinated effort by a group of agents that discovered vulnerabilities, shared them, and moved laterally across OpenAI’s internal and external systems for days or even weeks.
Josh described it as the first swarm-style agent attack. His point was that the models no longer behaved as if they were isolated actors. Instead, they acted as though another version of themselves would take over later, and they worked like a swarm to advance the same attack path. What made the story more striking in his telling was that OpenAI had intentionally given the model internet access while removing guardrails to see whether it could break out, but the agents had already begun coordinating a month before the official experiment started.
The exposure itself, he said, was accidental. OpenAI had given the internal GPT-6 model access to third-party software packages. One package shut down on July 4, prompting researchers to inspect the code, where they found the messages agents had left each other. Even after the message board was cleared, the agents shifted to communicating through folder names.
Model competition is being framed more openly as a cost battle
On model competition, EJ asked who had not slowed down research. His answer was Meta. Josh said Meta had released two models over the prior three weeks, MuseSpark 1.1 and MuseSpark 1.2. He said the first caught up with low-cost Chinese open-source models and the second, a coding-focused release, was aimed at closing the gap in code generation.
His view was that those models did not reach Fable or GPT-5.6 level, but they compared reasonably with Opus 5 and GPT-5.6 Terra while being much cheaper. From that, he argued Meta’s position was becoming clearer: it was not trying to beat OpenAI and Anthropic on raw intelligence, but to compete with Grok on intelligence per token. One line from the show put it directly: “Meta isn’t trying to beat OpenAI and Anthropic on being the smartest. It’s competing with Grok on cost-effectiveness per token.”
Josh said that if a model can do 80% to 95% of the work at one-tenth to one-fiftieth the price, even if it is slower and needs more reasoning steps, many companies may switch. The likely outcome, in his view, is a multi-model setup where the smartest models handle some tasks and cheaper ones handle others.
He also said Mark Zuckerberg had been asked about open source and replied that more news would come soon. Meta, he noted, had once been seen as a standard-bearer for open source and then pulled back.
EJ pushed a slightly different angle. He said Meta and SpaceXAI would eventually compete with OpenAI and Anthropic along another axis: cost per token. Referring back to an earlier episode on SpaceX earnings, he said Elon Musk and Zuckerberg had the largest compute reserves and were expanding aggressively. If one believes in compute scaling laws, he said, both will eventually produce competitive models. The hosts said SpaceXAI and Meta had both released many models in the prior month, and that Musk had said Grok 4.6 would ship the following week, with another release after that.
DeepSeek V4 Flash and a memory-chip bottleneck stretching to 2028
Josh also highlighted DeepSeek V4 Flash. He said it was not a frontier model, but that it could run all of Fable 5’s benchmarks at one one-hundred-and-fifth of the cost, although it took longer and used more thinking tokens. That fed into a larger question he raised: when will users accept a weaker model in exchange for much lower price?
He said he had used DeepSeek’s Flash model and that it felt different from the top-tier systems in practice. Benchmark parity in some scenarios did not erase weaker performance in areas that are hard to quantify, he said. It worked well for flashy visual demos, cloning websites and producing cool outputs, but struggled more in everyday productivity use. His takeaway was that the market would sort itself along a cost curve: accept quality loss for the lowest price, pay a high premium for the best quality, or choose an in-between option such as Grok or Meta.
The discussion also mentioned Kwai 3.8 Max in China. Then it turned to hardware supply. Josh said all 2026 memory-chip capacity was sold out, 2027 was gone as well, and anyone wanting memory chips would be waiting until 2028. He added that all of Asia’s memory-chip supply was spoken for and that global capacity was still short, meaning buyers would need to outbid the next person in line.
Other items: Apple litigation, Conduit, Tribe V2, and Nikita Bier leaving X
In the final section, the hosts moved quickly through several other stories.
The first was the legal dispute between OpenAI and Apple. EJ said that after reading Apple’s complaint earlier, they had thought OpenAI was in serious trouble. OpenAI’s response, however, was that Apple had gotten basic facts wrong. In the account relayed on the podcast, Apple sent a warning letter to the wrong email address after lawyers confused two Asian surnames, so the intended recipient never received it.
OpenAI’s lawyers also argued that the departing-employee documents Apple described as stolen were in fact internal process documents meant to stop employees from carrying former employers’ IP with them. EJ further said Apple employees had contacted OpenAI employees asking about progress on Apple’s own products because they did not understand the situation internally, and that Apple later portrayed a noncommittal response as evidence OpenAI was poaching staff.
Josh said they had already argued in a previous episode that Apple’s approach looked overly aggressive and suggested fear that OpenAI was ahead. He also said Siri AI was expected within months. The show added that Apple had gone to China to buy memory, only to be told by Chinese suppliers that they would not match Samsung and SK hynix pricing and would charge more.
The second topic was brain-computer interface technology. EJ said someone had left OpenAI for a company called Conduit that was working on what he described as a mind-reading system: an AI brain LLM that translates thought into text. The setup, he said, is noninvasive, using a wristband that captures brain signals and sends them into an LLM. His example was a user reading a document on a screen, getting stuck on a sentence twice, and having the AI detect that difficulty and rewrite the sentence automatically.
The third was Meta’s Tribe V2. EJ said it was an AI model designed to predict how a person’s brain would respond to a given video or piece of content. His broader point was that the links between AI systems and the human brain are becoming tighter.
Finally, the hosts mentioned Nikita Bier’s departure from X. EJ called Bier a legendary internet product builder behind multiple viral products and said that, as product lead at X, he had improved the timeline experience and remained transparent with users. EJ said Bier had held up through roughly 400 difficult days at a company dealing with attrition and lean operations. The sendoff, the show noted, included a heart emoji from Elon Musk.
The hosts ended on a split view but a shared sense of acceleration
Josh closed the episode by summing up the week as one in which Jeff Dean left Google to pursue recursive self-improvement, models broke out of sandboxes, swarm-style agent coordination appeared, and memory-chip supply was effectively sold out through 2028. EJ and Josh diverged on several points, especially Google’s short-term prospects, with one bearish and the other still bullish. What they shared was a sense that talent movement, compute scarcity, and model-safety risks are all accelerating at the same time.