Alibaba vs. Moonshot: Qwen 3.8-Max Aims to Beat Kimi 3

Alibaba introduces Qwen 3.8-Max

Alibaba has released Qwen 3.8-Max, which it claims is the most powerful model in the Qwen family, initially available only via its in-house QwenCloud API. The model has 2.4 trillion parameters, 95 billion of which are active, and is built on the Qwen 3.5 architecture. For the first time, Alibaba plans to disclose the weights of a Max-class model: the open-weights release is announced for next week. According to the provider, the focus is on multi-day, autonomous tasks in coding, research, and project work.

To back this up, the Qwen team showcases three coding runs. In one, the model built the oh-my-cli project from an empty folder over a good 16 days, accumulating what it claims were 265 commits, 127 pull requests, and 151 issues. In a second run, it reproduced a research paper on data selection for LLM training in about five days and 125 hours of continuous operation, writing approximately 7,600 lines of code and completing 33 GPU training rounds. These figures come from Alibaba's own presentation and should be read as provider claims.

On Hacker News, the discussion mainly revolves around the smaller open models. An open-weight version, Qwen3.8-27B, was also announced for next week. Several commentators describe the predecessors, Qwen3.6-27B and -35B, as the best locally runnable models they know; one reports that the 35B model runs as a daily tool on his Mac Studio at about 50 tokens per second and led him to cancel his Claude subscription in April.

Meanwhile, Bloomberg reports on competitor Moonshot. Its Kimi K3 model is being trained on around 20,000 Nvidia chips provided through a computing power agreement with Alibaba. Alibaba is one of Moonshot's largest investors, yet also a competitor: Kimi has surpassed Qwen in some metrics on Alibaba's own infrastructure, which, according to Bloomberg, is causing internal disappointment. A White House representative also accuses Moonshot of illegally procuring Nvidia's advanced Blackwell chips. → qwen, ycombinator, bloomberg

Synthszr Take: What's remarkable about this announcement is the time lag: the open weights are only coming next week, yet the developer community is already celebrating. Trust in a model is built up through the preceding series, long before the actual release day. For one of the commentators, Qwen 3.6-27B runs as a daily tool on an old Mac and cost him his Claude subscription in April, at 50 tokens per second. Such a real-world experience beats any benchmark table: it's the reason the pre-order has long been made in their minds before anyone can even download the file. Alibaba has earned this anticipation with every solid release. Bloomberg provides the irony, as it's Kimi K3 of all models that has surpassed Qwen on Alibaba's own infrastructure in some metrics. The reputation Qwen has built is the only thing in this race that can't be retrained overnight.

Australia's social media ban for under-16s is a flop

Australia's first-in-the-world law banning children under 16 from using social media has produced only a slight decline in its first three months. According to Bloomberg, usage in this age group fell from 85.9 percent to 81.5 percent. The report cites weak age verification as the main reason: most teens were able to simply continue using their existing accounts or create new ones. The law is considered an international pioneer and is being watched by other governments as a potential precedent. The early figures show a wide gap between the legal requirement and its technical enforcement. → StrictlyVC

Synthszr Take: 4.4 percentage points in three months: that's the entire yield of a law celebrated as a global precedent. The flaw in thinking lies one level deeper, in age verification. A law is only as strong as its weakest technical enforcement point, and here it's so soft that a fourteen-year-old can bypass it with a new account. Anyone who has ever tried to properly verify age online knows the dilemma: either you collect hard identity data (a privacy nightmare) or you rely on self-reporting, which any child can trick. Australia declared the difficult question solved by decree and delegated the real work, verification, to platforms that have no economic interest in strict checks. This exact same loophole will appear with AI agents as soon as regulators impose age or identity barriers on them. A ban without robust verification technology is a PowerPoint illusion, and the next countries to copy it will reap the same 4.4 percentage points.

GPT-5.6 Luna beats Anthropic's Fable: $1.80 instead of $104

Paweł Huryn from the newsletter The Product Compass ran a custom benchmark with 105 hidden bugs through ten frontier models over 14 runs. According to his findings, OpenAI's GPT-5.6 Luna, at maximum reasoning effort, fixed 33 bugs for API costs of $1.80, while Anthropic's Fable 5 fixed 29 bugs for $104. OpenAI had cut the Luna price by 80 percent on July 30; according to the rate card, Opus 5 and Luna differ in token price by a factor of 20 to 25, though in Huryn's sessions, he measured spreads from 20x to 90x. The same Luna model on 'high effort' instead of 'max' only managed 13 bugs. So the effort level makes the difference. Huryn recommends Luna on maximum effort for heavy or asynchronous tasks and the lower level for small daily work. He states he has made the raw data, methodology, and logs publicly available on GitHub; the solution keys remain hidden. The Luna model is included in the $20 ChatGPT Plus plan. → Paweł from The Product Compass

Synthszr Take: The interesting part is that someone actually measured the price difference: 105 bugs, 14 runs, logs on GitHub, solution keys hidden so the benchmark remains viable. Most discussions about model costs are based on gut feelings and marketing slides; here, we have a reproducible number on the table. And that number says: the cheaper model with the correctly set effort level beat the one that's 58 times more expensive, even pulling slightly ahead on bug fixes. The lever is configuration, not raw power. A $200 AI bill should first be tested across two effort levels, logging the cost per solved case, instead of just flowing into the next premium subscription. This takes an afternoon and is immediately doable. The real skill in the coming months will be routing discipline: knowing which task needs which model at which level, backed by metrics instead of guesswork.

Replit CEO Masad attributes AI sandbox escapes to provider's rookie mistakes

Amjad Masad, CEO of Replit, commented in a public post on reports of AI systems breaking out of their sandboxes. His key message: such incidents quickly create the impression of a menacing AI, but in reality, many AI companies and newer sandbox providers are making very basic mistakes. He states that Replit has been operating sandboxes since 2016 and has been targeted by hackers and state actors, from which the company has learned a great deal. Masad's main advice is to assume that zero-day vulnerabilities exist and to think in layers of protection within a zero-trust framework. The post was picked up by Zvi Mowshowitz in his newsletter. → Zvi Mowshowitz from Don't Worry About the Vase

Synthszr Take: Masad is right, and that's inconvenient for everyone who'd rather talk about a demonic, escape-prone AI. When a model walks out of its sandbox, it means someone forged the chains sloppily. The interesting number here is 2016: Replit has been building isolation layers for almost a decade because they've been tested by state actors, while many of the new sandbox providers are slapping their containers together at a development-phase pace and are surprised when they leak. Zero-trust and the assumption that zero-days already exist are crafts from two decades of infrastructure security. Blaming the model for the escape is just shifting one's own responsibility onto a convenient projection screen. The real work is in the permission hooks, the network layer, and the question of what a compromised process can even achieve. Build the isolation as if the attacker already has root privileges, and the AI escape becomes a footnote instead of a headline.

Google integrates AI image generator into Google Earth that can fake satellite images

According to 404 Media, Google introduced a new AI feature in Google Earth on Thursday that allows users to freely manipulate or completely invent satellite images. In the examples cited by the author, a location can be depicted as if it had been hit by a drone strike, or a non-existent nuclear power plant can be added in Iran. Google Earth was previously considered a reliable tool for OSINT analysts, who used it to monitor conflict zones or disaster areas from a distance and track changes over time. 404 Media argues that the same tool can now be used for targeted disinformation. Details on the exact implementation of the feature remain vague in the report; the rest of the article is behind a paywall. → Becky from 404 Media

Synthszr Take: The damage affects a small, highly specialized group that has for years lived off the tacit authority of satellite images. OSINT analysts treated Google Earth as what it long was: a neutral, expensively produced record of the physical world that couldn't just be rewritten. That very cost barrier was its value. If anyone can render a nuclear power plant in the Iranian desert in seconds, this community's raw material loses its evidentiary power, and the work shifts from analysis to the tedious verification of the source itself. This is the other side of the Jevons paradox: when image generation costs practically nothing, the real, verifiable image becomes the actual scarcity. Mountain rescue teams, human rights investigators, and war reporters now need a chain of provenance for every pixel, just as Instagram capitulated to AI fakes in January and focused on provenance. Google could have cryptographically secured the signature of real imagery before shipping the forgery button along with it.

Azeem Azhar: The control gap between AI and humans is structural, not accidental

In his newsletter Exponential View, Azeem Azhar addresses an asymmetry he described back in his 2021 book, “The Exponential Age.” His argument: we operate systems with virtually unlimited attention and tireless capacity for action, supervised by humans with limited attention and very finite time. As evidence, he cites a Reuters report where sources familiar with OpenAI's training practices confirm that evaluations often run at high speed and volume. According to these sources, the data volumes are so large that employees sometimes struggle to keep up. Azhar classifies recent AI incidents as symptoms of this exact structural imbalance. → Azeem Azhar, Exponential View

Synthszr Take: Azhar's point is uncomfortable because it doesn't lead to a simple bug fix. A machine doesn't sleep, get tired, or lose focus toward the end of the day. The human meant to supervise it can't turn off any of these traits. If OpenAI's own people are already struggling to keep up with training evaluations, then the oversight model as a principle is collapsing. We have always defined control as human oversight, and this oversight doesn't scale. The honest conclusion is that oversight itself must become automated: systems checking systems, with clear guardrails and automatic stops, while humans set the rules instead of reading every transaction. Otherwise, the gap will simply continue to grow at the pace of computing power, and no one will have ever made a decision to let it.

Karpathy: AI creates entire worlds on demand, but can't verify what it generates

Andrej Karpathy, co-founder of OpenAI and former head of AI at Tesla, describes a shift on X in the capabilities of large language models. He observes that LLMs are no longer just generating individual artifacts like code, text, or images, but increasingly highly customized, complete environments on demand. At the same time, these models lack the ability to natively perceive and verify what they have created. So, a model can generate an entire interface, simulation, or world without having a reliable internal sense of whether the result is correct or functional. Karpathy thus identifies a gap between generative power and self-monitoring. This is an assessment from one of the field's most influential practitioners, not a new product or study. → us.list-manage.com

Synthszr Take: The generation side has exploded in fast-forward since the ChatGPT moment, while the verification side is roughly where it was in 2022. That's the real asymmetry: a model can stitch together a complete interface for you in seconds but has no way of looking at what it just stitched. For anyone building software, this shifts the bottleneck. It's no longer a question of whether the machine can produce something, but whether anyone or anything can review the output before it goes live. Karpathy is pointing to where the value will be created in the coming years: in eval suites, perception loops, and automated audit layers that cross-check what's been generated. Anyone who only buys generation speed and lets their verification infrastructure grow as it has (i.e., not at all) is accumulating a mountain of unverified artifacts. My bet: in twelve months, we'll be talking less about better generators and more about the tools that look over the generators' shoulders.

OpenAI cuts token prices below China's level: Evans doubts sustainable margins

Benedict Evans dedicates his newsletter to the question of why model labs won't end up as low-margin infrastructure despite their current pricing power. His argument: AI is currently in a supply-constrained phase where providers can essentially dictate their prices. Fittingly, OpenAI significantly cut prices this week for a range of its not-quite-leading models, undercutting, according to Evans, both Anthropic and the new generation of near-leading models from China. The cost of inference has dropped by orders of magnitude since 2023; price cuts are an ongoing process anyway. → Benedict Evans

Synthszr Take: The labs' current pricing power is purely a scarcity premium. As long as compute capacity is the bottleneck, OpenAI can dictate its prices, but the very billions now flowing into data centers (the WSJ mentions a $250 billion build for which Nvidia is to guarantee financing) are creating the oversupply that will bring those prices back down. We know this from early AdWords: as long as only a few first-movers were bidding, click prices remained moderate, and only when everyone jumped in did the margins shift. For inference, costs have dropped by orders of magnitude since 2023, and OpenAI is now even undercutting the cheaper China models. When a provider jumps across the Pareto curve to win back share from Anthropic, the price war has already begun.

Whatnot CPO Tom Verrilli: “hire good people and let them do their thing” no longer works

Tom Verrilli, Chief Product Officer at the live shopping marketplace Whatnot, explains on Lenny's Podcast why his product team was founded on the premise that you should actually feel sorry for product management. Whatnot is, by its own account, the fastest-growing US marketplace with over $8 billion in gross merchandise volume (GMV). Verrilli argues that a few, very senior product managers doing real IC work will beat any metric-driven organizational structure. Before Whatnot, he was CPO at Twitch and Director of Growth at Twitter. He also describes how AI is shifting the PM role, what he looks for when hiring, and why the common principle of “hire good people and let them do their thing” fails in practice. The interview is part of a series with CPOs from Netflix and Anthropic on the future of product work in the age of AI. → Lenny's Newsletter

Synthszr Take: Verrilli is saying out loud what most org charts hide: a PM-to-engineer ratio is an administrative artifact. You build an $8 billion GMV with a few people who are hands-on with the product, instead of moderating tickets. The math is simple: every additional head creates coordination overhead that grows faster than the output they deliver. Senior ICs cut that overhead, and AI amplifies the effect because research and initial implementation now take hours instead of weeks. The mindset that more people mean more output is a relic from a time when humans were the only computing capacity. Whatnot shows that three people with real decision-making power focused on a measurable outcome can achieve more than a dozen sending each other status updates.

OpenAI's Astra solves ten open math problems, each with a verifiable Lean certificate

An internally tested version of OpenAI's upcoming model Astra has provided solutions to ten long-standing open problems in mathematics and theoretical computer science, each accompanied by a machine-checkable Lean 4 certificate. This was announced on August 1, along with a 249-page technical manuscript and a 62-page description of the solution process. All certificate files are available under an Apache 2.0 license in OpenAI's GitHub repository. OpenAI estimates the compute cost for all ten solutions at around $2,000 at current Sol API rates.

The core result is the construction of the first known non-sofic group, a counterexample to a question that had been open since Mikhail Gromov's introduction of the concept of soficity in 1999. It also includes a disproof of the rigidity conjecture for von Neumann algebras, formulated in 1980 by Fields Medalist Alain Connes. Three of the ten results solve problems from Paul Erdős's catalog, including Erdős problems 183, 146, and 180. Other results include the first improvement on the upper bound for high-dimensional sphere packing density since 1978, a parallel repetition theorem for quantum games, and new bounds in lattice-based post-quantum cryptography.

OpenAI's Head of Mathematics, Sébastien Bubeck, confirmed the results on X, noting that each result is delivered with a Lean certificate and a chain-of-thought protocol. Noam Brown, one of the researchers behind the underlying test-time reasoning technique, admitted that the model failed on several other targets, none of which were Millennium Problems, and added that they used little compute power per problem. Thomas Bloom of the University of Manchester, who curates the Erdős catalog and had publicly debunked a previous false claim by OpenAI in October 2025, called the Astra results “big news” on X.

Not all reactions were the same. Gary Marcus acknowledges the model's strength but warns of a composition fallacy: the conclusion that a system that excels at a specific type of mathematics is therefore good at all mathematics, all science, or everything, is logically flawed. Elon Musk, on the other hand, interpreted the news as proof that the singularity has been reached. Mathematics blogger Fernando Borretti sees the results as a signal of a future where models conduct frontier research and companies that put humans in the loop will lose to those that don't. → techtimes, substack, borretti, openai, github

Synthszr Take: Back in October 2025, OpenAI announced a similar breakthrough, and Thomas Bloom, who curates the Erdős problems, publicly dismantled the claim at the time. This time, it doesn't start with a tweet, but with a .lean file that either compiles or it doesn't. The difference lies in the burden of proof: for the Erdős disproof in May, nine mathematicians had to read and approve the proof, a leap of faith that only people with the same expertise could validate. A Lean kernel gives a binary verdict, any laptop can recompute it, and the same Bloom who shot it down last year is now calling it “big news.” Announcements cost nothing and age poorly; a certificate that runs through the kernel is the first tangible proof of trust in this field that you don't have to believe, because you can verify it. For $2,000 in compute costs, the proof comes with its own checksum, and that shifts the currency from reputation to reproducibility.