AIs are better doctors, superheroes, musicians — let’s celebrate our imperfection

Orchestrated Frontier Models Likely Better Than Our General Practitioners

Microsoft has published a study in the journal Science showing that its in-house diagnostic system, MAI-DxO, correctly identifies complex medical cases more than twice as often as general practitioners. In 304 difficult cases from the New England Journal of Medicine, the system achieved an accuracy of 83 percent. General practitioners and internists reached 36 percent with reference materials, and around 22 percent without. For more typical emergency room cases, taken from the MIMIC-IV-ED database of the Beth Israel Deaconess Medical Center in Boston, the ratio was 80 to 47 percent. The paper, titled 'Sequential Diagnosis with Language Models,' had already been released as a non-peer-reviewed preprint last year. Microsoft AI chief Mustafa Suleyman calls the result a step towards 'medical superintelligence'.

MAI-DxO works sequentially. The system initially receives only the chief complaint and decides for itself which questions to ask and which tests to order. Multiple AI agents take on roles as in a medical panel: they propose diagnoses, select tests, question assumptions, and weigh costs while adjusting the Differentialdiagnose over several rounds. The researchers argue that this can compensate for cognitive biases like anchoring on an initial suspected diagnosis. For the tests, Microsoft developed its own benchmark called SDBench because certification exams like the US USMLE provide all information upfront and thus do not reflect clinical practice. The company plans to release parts of it, including the evaluation and cost estimation code, as open source.

The system achieved the best results with OpenAI’s GPT-5 and GPT-5.5. Models from Google, Anthropic, Meta, and DeepSeek were also tested. The study data shows that a large part of the performance comes from the underlying model. GPT-5 alone achieved 78.9 percent in complex cases at an estimated cost of around $8,400 per case. The orchestration reduced these costs by about $3,000 and increased accuracy by about four points. With GPT-5.5, the score for complex cases rose from 72 to nearly 87 percent, while for emergency room cases, both variants were around 85 percent. Weaker models benefited the most, such as Gemini 2.5 Flash, which climbed from 46.4 to almost 69 percent. Microsoft attributes the diminishing gains to the models approaching their theoretical maximum. Newer models like GPT-6 Astra or Opus 5.5 are not part of the study.

The limitations are significant. Testing was conducted exclusively on historical cases. Specialists and medical teams who treat such cases in practice were not part of the comparison, and the authors themselves expect the gap to be smaller. The claim that the system reaches the correct diagnosis at one-third to half the cost is based on estimated costs in simulations, according to Microsoft. Medical professionals emphasize that such systems are intended as assistants and point to unresolved liability issues. → TechCrunch, The Verge, Gizmodo, Microsoft AI

Synthszr Take: Four more points of accuracy and $3,000 less per case—that’s the entire contribution of orchestration beyond GPT-5. The model delivers the rest, and that belongs to OpenAI. Microsoft’s capital lies in SDBench, a benchmark that evaluates questions and test orders and defines what 'clinically validatable' means. Any provider wanting to enter clinics must measure themselves against this yardstick, which Microsoft built itself. Suleyman’s superintelligence is marketing; the standard is the product.

OpenAI to Raise Another $30 Billion Before IPO

OpenAI is in talks for another private funding round of up to $30 billion at a pre-money valuation of $1.25 trillion, according to CNBC. Sources told CNBC that Coatue and investors from the United Arab Emirates are expected to lead the round. A month ago, Bloomberg reported on a round of over $30 billion at a $1.4 trillion valuation. Both media outlets state that the terms are still being negotiated and could change. According to CNBC, the money from the Emirates could be linked to an agreement for additional computing capacity for OpenAI. → TechCrunch

Synthszr Take: The price of a private round is set at the table with a handful of buyers, without a prospectus and without quarterly figures. How soft it is can be seen in the month between two reports: $1.4 trillion according to Bloomberg, $1.25 trillion according to CNBC, and no published business figures have changed. Money from the Emirates tied to computing capacity makes the valuation even more elastic: the buyer is partly paying with their own goods. As long as the IPO documents remain confidential, the trillion-dollar figure remains a matter of negotiation.

OpenAI’s Steinberger Burns Through $10 Million in Tokens Per Month Alone

Peter Steinberger, developer of the open-source agent OpenClaw and now employed by OpenAI, spent an estimated $10 million on AI tokens last month. This was reported by the newsletter AI Secret. In May, Steinberger had published a monthly bill of $1.3 million, meaning his spending has increased about eightfold since then. His employer, OpenAI, covers the costs. At a Cloudflare conference, the moderator noted that Steinberger has a 'virtually unlimited token budget,' which not everyone in the room has. → AI Secret

Synthszr Take: Ten million dollars in tokens per month for a single developer is a financing issue. Extrapolated, that’s $120 million a year that will land in the funding round OpenAI needs to raise before Anthropic’s IPO. Every dollar burned increases the capital requirement, and every round increases the pressure to show growth with even more compute. Steinberger is dependent on a budget that no one outside of OpenAI would pay for, not even the audience at Cloudflare. Compute discipline would shrink the very story that is raising the money.

Attackers Exploited PHP2Shell Flaws Four Days Before Patch

Software vendors, security firms, and authorities have issued warnings, detection rules, and countermeasures for two critical PHP vulnerabilities known as PHP2Shell, according to SecurityWeek. Many vendors are still checking if their own products are affected. The vulnerabilities, CVE-2026-109618 and CVE-2026-109619, allow unauthenticated attackers to execute remote code. They were discovered in August by Wiz, OpenAI, and Calif. PHP released the fixes on October 21, a day earlier than planned, because the CVE identifiers were already circulating online. → SecurityWeek

Synthszr Take: Half of all PHP systems will never get this patch, estimates Cisco Talos, and this half belongs to small and medium-sized businesses, clubs, and shops on hosting packages without their own admin. Google Cloud deployed the updates automatically, and cPanel and Plesk even provided builds for long-deprecated PHP versions. This means the hoster decides who is secure. For the market, this means: automatic patching becomes a selling point, and self-managed servers lose their last justification: cheap was yesterday.

Unitree Stages New Robots as Superheroes

Unitree Robotics has introduced a humanoid robot called the H2-X, which the company presents as a 'Guardian' for emergencies. According to Interesting Engineering, the promotional video shows the machine disarming an attacker with a knife, smashing a car door, lifting a vehicle, and pulling a pedestrian away from an oncoming car. Unitree has not disclosed whether the scenes were autonomous, remote-controlled, or pre-choreographed. The H2-X is built on the existing H2 platform, which, according to the manufacturer, is about 1.8 meters tall, weighs 70 kilograms, and has 31 Freiheitsgrade. It has not been confirmed whether the new model retains these specifications. → Techpresso

Synthszr Take: The video works for Unitree even if no scene was autonomous. It sets the bar for the entire industry before a single field test exists. Every competitor now has to explain why their machine doesn’t disarm someone with a knife, even though no one knows if the H2-X can. The warning at the end to maintain a distance of two to three meters reveals the state of the art more accurately than the rescue scene. Nevertheless, procurers, insurers, and regulators are calibrating to the image.

Musk’s Robots Are to Educate Themselves

According to Elon Musk, Tesla needs to build between 10,000 and 20,000 Optimus humanoid robots so they can learn tasks through practice. The Tesla CEO said this during the third-quarter earnings call on October 21. The robots are supposed to practice various activities in a so-called Optimus Academy. The data generated through Self-Play will then be used to improve the controlling AI. Musk compares the process to raising a toddler, who is initially wobbly and learns through play. → Techpresso

Synthszr Take: The first 20,000 Optimus units are Tesla’s data factory. With cars, customers pay for vehicles that incidentally train the driving software. The robots lack this customer until the end of 2027; Tesla is self-financing the fleet and for now only gets data in return. The capital is flowing into training halls and camera helmets for trainers; revenue will come in two years at the earliest. For competitors, the decisive factor will therefore be whose robots are the first to deliver data in real-world operation while already being paid for.

AI-Native Companies Are Becoming Flatter: Executives Are Building Again

AI-native companies like Block, Lovable, Every, and Cognizant are reorganizing their management structures around AI agents, reports the Substack newsletter Growthwise Teams. The author points to a Gartner forecast from late 2024: by the end of 2026, one in five organizations is expected to flatten its hierarchy with the help of AI, eliminating more than half of its middle management positions. The evidence for this comes mainly from self-descriptions from the companies. Block CEO Jack Dorsey and Roelof Botha describe the restructuring of their company in the text 'From Hierarchy to Intelligence.' Growth manager Elena Verna reports on a year at Lovable, and Every CEO Dan Shipper describes in 'After Automation' how his company operates after extensive automation. → Substack

Synthszr Take: One in five organizations by the end of 2026 is a measurable bet, and so far, the evidence comes from companies whose business is AI and who are reporting on themselves. The real test is happening elsewhere: in corporations where middle management primarily produces upward reporting. Agents write these reports faster than any team leader. The position will only disappear when the board stops ordering the report. That will determine whether Gartner is counting or just being quoted.

What’s Left for Us: Imperfection

Musicians are increasingly using off-key notes, unconventional melodies against odd time signatures, and new genre mixes like Celtic EDM to distinguish themselves from AI-generated music. The starting point is the ability of AI systems to create songs in the style of Taylor Swift, Fleetwood Mac, or Frank Sinatra. In parts of the scene, imperfections are now considered proof of human authorship. David Das, Vice President of the Society of Composers and Lyricists, points out that music could be produced perfectly even before AI, with perfectly timed drums and pitch-corrected vocals. In his opinion, AI is pushing this smoothness even further.

Cecil Coleman, drummer for the Australian band Body Type, observes a trend towards deliberately 'messy' playing. There is a broad consensus among musicians that AI music sounds overproduced and too polished. Many artists she admires are therefore choosing an earthy, raw sound to set themselves apart. Das himself now records without a click track and without samples. He describes this as a psychological rebellion against flawless productions: music should feel human again.

Experts cite technical reasons for the typical smoothness of AI music. Robert Rowe, a professor of music innovation at New York University, explains that the models are designed to predict the next element rather than planning a piece as a whole. From phrase to phrase, the result sounds convincing, but over an entire song, it often doesn’t make sense, repeats strangely, or too closely resembles existing works. Carmine-Emanuele Cella, a composer and co-director of the Center for New Music and Audio Technologies at UC Berkeley, attributes this to Autophagie: when models are trained on machine-generated data, a feedback loop is created. This results in patterns that can be recognized similarly to AI-generated text.

According to Cella, typical tell-tale signs include excessive repetition in lyrics, choruses, and harmonic progressions. Audio engineer Liam Weiland mentions a high-frequency noise as another characteristic. In his estimation, average listeners usually don’t notice such features or don’t question their cause. Weiland expects these indicators to disappear soon with better models. A historical parallel is drawn to photography in the 19th century: it triggered existential fears among painters and led to Impressionism and abstract art. → The Deep View, The Deep View

Synthszr Take: An off-key note is only a signal of authenticity as long as it remains expensive. AI makes perfection free, so the market rewards proof of human origin, and musicians deliver it: without a click track and without samples. The catch lies in the training loop: every imperfection becomes the next pattern for the models to learn (according to Weiland, the high-frequency noise will soon disappear). Labels and platforms need proof of origin in the metadata because the sound itself will only carry it for one model generation at most.

Zuckerberg’s Control Cemented: Meta Funds AI with Non-Voting Shares

Meta is introducing a new class of non-voting stock to finance employee compensation, acquisitions, and capital raising without Mark Zuckerberg losing his majority voting power. The Class C shares will be technically issued as a special dividend on a one-for-one basis to all shareholders. Meta announced the move on Wednesday along with its third-quarter 2026 results. This followed months of sometimes tense negotiations between Zuckerberg and a board committee that included DoorDash CEO Tony Xu, former ambassador Robert Kimmitt, and former Microsoft executive Charles Songhurst. The committee even considered resigning at one point. Xu disclosed his personal friendship with Zuckerberg but stated that it did not affect his judgment.

According to the SEC filing, Zuckerberg had three goals: maintaining his voting control, fulfilling his commitments to the Chan Zuckerberg Initiative, and preparing for a potential California wealth tax. This tax, Proposition 40, will be voted on November 3. In parallel, Meta had decided at the end of 2025 to raise external capital for its AI expansion. Zuckerberg holds about 13.4 percent of the shares but controls 60.6 percent of the votes through the Dual-Class-Aktienstruktur. His Class B shares carry ten votes per share and convert to simple Class A shares upon sale.

Previously, Meta offset the dilution from newly issued Class A shares with buybacks, which increased Zuckerberg’s voting share from 53.7 percent in 2013 to 60.6 percent. In 2025, Meta bought back $26.3 billion worth of Class A shares, but has not bought back any in the last four quarters. In the first nine months of this year, stock-based compensation totaled $21 billion. Zuckerberg is allowed to sell 20 percent of the new Class C shares immediately. The remaining 80 percent are tied to his voting shares and can only be sold if he simultaneously sells or converts an equal number of Class B shares.

The quarterly figures reveal the financial background. Revenue increased by 26 percent to $64.6 billion, while capital expenditures reached $34.9 billion. Free cash flow dropped to minus $2.1 billion, down from plus $10.6 billion in the same quarter last year. For 2026, Meta is raising the lower end of its capital expenditure guidance from $130 billion to $135 billion and announces continued high spending for 2027. For the fourth quarter, the company expects revenue of $73 billion to $76 billion.

The results were burdened by a settlement of about $18 billion with 52 attorneys general regarding the mental health of adolescents. In the third quarter, this amounted to a $9.89 billion charge. Earnings per share were $3.83; without the special charge, it would have been $7.40. The stock nevertheless rose four to five percent in after-hours trading. → Variety, The Information, Fortune, Constellation Research, Shacknews, Coinpedia Fintech News

Synthszr Take: Four quarters without buybacks, free cash flow at minus $2.1 billion. The old mechanism Meta used to maintain Zuckerberg’s voting share has fallen victim to the AI expansion. The non-voting Class C shares are its replacement. Employees and acquisition targets will now be paid in paper that has no say, and Zuckerberg gets to sell 20 percent of it immediately, conveniently before the wealth tax vote on November 3. Shareholders are funding $135 billion in capex and getting less vote per dollar in return.