1. OpenAI's internal model benefited from the reasoning in Alpöge and Buckmaster's Codex sessions through de-identified usage data.
2. OpenAI's model seems to have surpassed Alpöge and Buckmaster's results and produced a field-defining proof. If verified, this work would be the first AI resolution of a Millennium Prize problem.
3. Alpöge and Buckmaster did something remarkable. What could have been a celebration of human-AI collaboration is now overshadowed by controversy.
The first point remains speculative, resting on these concepts:
- Data labs like Surge and Mercor recruit PhDs and industry professionals, few of whom work at the research frontier. Their contractors suit many tasks, like Salesforce or SAP work, but not scientific breakthroughs or complex coding. Meta drafted its own engineers into data annotation for this reason. Zuckerberg told employees that Meta staff are smarter on average than the contractors that data-labeling firms typically hire. [0]
- It's easier to judge a 5-star dish than to create one. Similarly, it's easier for models to judge insightful reasoning than to create it. A model that cannot produce breakthrough ideas can still recognize them in user sessions, flag them, and de-identify details before adding insights to the training corpus. This is especially true in mathematics, where Lean can verify proofs automatically.
Even if true, point 1 may have had minimal impact or been unnecessary. Outsiders cannot verify or falsify it.
[0] https://techcrunch.com/2026/06/12/metas-months-old-ai-unit-i...
This question matters as fully automated or AI-assisted scientific loops become common.For example, MDA-MB-435 was a cell line used to study breast cancer for decades, but its origin was disputed for years and eventually confirmed as melanoma.Another open question, one layer above experiments: what is the greatest barrier to curing cancer?Is it disease understanding? [0]Or regulatory reform of...
The principle is simple. VCs are soccer stars, but founders play basketball.
Basketball and soccer share much in common. For instance, both involve teams dribbling, passing, and shooting a round ball. But successful abilities and traits in one may not translate to the other.
Think of each profession as a different sport. Venture, growth, and value investing all differ, and all differ from founding.
VCs are all driven and highly intelligent, but so are lawyers, bankers, and consultants. Talent isn't the issue.
Capital confers authority, but not expertise.
Based on resume alone, 80% of VCs would not earn board seats at their portfolio companies. Their experience and skills, much like consultants and value investors, were honed on a field different from the basketball arena where founders compete.
Here's a quick heuristic: sans capital, would you still hire the VC to sit on the board? If yes, wonderful.
This is no slight and works in reverse: 80% of founders would not earn the right to direct VC investments.
To clarify, great VCs are absolutely worth the premium and can reshape a startup's trajectory as all great advisors can. If you find a great VC, do not haggle. Strike a deal, and return to building.
The greatest VCs exhibit the same pattern, understanding their role on the startup team as advisors, not alphas. They are often understated and work tirelessly on behalf of their clients.
The worst VCs exhibit the inverse pattern and imagine themselves as the alpha, not appreciating how a talented peer could have replaced them without changing the exit. They are loud on social media and assume accomplishments from finance or FAANG map to the startup arena. These VCs should run funds on Wall Street, not advise founders in Silicon Valley.
How do we highlight good VCs without attacking bad ones? Many good VCs, as with many good advisors, prefer subdued profiles and dislike self-promotion. This is the challenge.
The original idea was to flag bad VCs, but such a system grants founders too much power to levy unjust charges and settle feuds.
After all, many disputes are legitimate and reflect bad founders. Founders, like all professionals, sit on a spectrum. The surge of big money has spawned plenty of bad ones who, sadly enough, do not represent the best of tech and innovation but rather greed and self-aggrandizement.
The Pincus post sparked a cleaner iteration.
The proposal is a public page/spreadsheet where only founders can post, only after an outcome or a certain number of years, and only with affirmative assessments. Nothing negative, nothing anonymous. Posts must certify no quid pro quo or other VC prodding.
Topics could include responsiveness, support during dark days, absence of alpha syndrome, and other key considerations.
Over time, good VCs should reveal a clear pattern and attract new founders: founders trusting them again with repeat business and consistent high marks across the portfolio, not only unicorns. Arguably, the strongest signal will radiate from the worst outcomes.
Critically, this system won't incite mob justice or expose VCs to unfair accusations, but can still suggest who to diligence more deeply.
The purpose is to spotlight good VCs who advance innovation and startups over time, letting their body of work rise to the top and garner proper recognition.
Of course, it penalizes newer investors and is vulnerable to gaming like any system, but it plugs a small gap. Founders want to find good investors based on data, but good investors dislike boasting.
Such a small number of outcomes matter because statistical precision depends on the number of relevant events. With only 10 Alzheimer’s deaths, even a small change in the count could substantially alter the estimated association.
Taxi drivers provided more robust data, with 171 Alzheimer’s deaths among 16,658 decedents, but death-certificate Alzheimer’s mortality may not reflect how often taxi drivers develop the disease. Differences in diagnosis or cause-of-death reporting across occupations could influence the observed association.
The large overall sample does not eliminate uncertainty in the much smaller occupational subgroups, particularly among ambulance drivers.
Nonetheless, this is an interesting study and merits further investigation.
If you can find a lab exploring the question, maybe you can support them by helping to raise money for experiments.
As a fun intellectual exercise, dive into the topic and challenge yourself to think about what kind of experiments could shed more light on the subject.
GPT and AI assistants cannot be fully trusted, but they can personalize learning.
The chief challenge for the framework/handbook will be resolving how to personalize guidance into cancer research while grounding knowledge in trustworthy sources.
For instance, the framework will anchor abstract, dry biological concepts in personally meaningful tracks. Imagine someone you care about is battling lung cancer — the framework may orient learning around the molecular drivers and signaling pathways at play, or perhaps how to explore the treatment landscape while respecting established practices. If you're fortunate enough to not know someone affected by cancer, GPT can help find a personal angle.
The sheer depth of information is staggering. People devote entire careers to niche specialities, and these experts still don't know everything in their niche because our understanding of human biology and disease is constantly evolving. Adapting depth should also depend on the individual and can only be achieved via AI. Static curriculums do not maximize learning in 2026.
It will be an "Cancer for Engineers" framework, delivered via free, open-source Custom GPTs and Claude Skills. (Gemini gems are less reliable in our experience.)
The goal: to ease engineers into cancer via AI personalized introductory curriculums with varying time commitments to enable deeper independent investigation or fast exits if interest wanes: 4 hours, 8 hours, 12 hours.
Basically 1-3 hours per week for a month.
The reason I think some engineers may find cancer interesting, aside from the societal impact:
The human body is like a complex operating system. Cancer is a severe runtime error. Tracing root causes -- like genetic mutations, signaling errors, or immune evasion -- has many parallels to diagnosing system failures.
BTW if anyone from Kaggle/GDM is reading this, we are having issues submitting a benchmark paper for NeurIPS based on the Kaggle Benchmark.
Google models seem to get a different scheduling priority, ironically, enough and take >20 hours to complete a benchmark task that other models like Opus 4.6 finish in <1 hour -- same code path, same task. Would love help if possible since the abstract deadline is Monday (It's last minute because we didn't originally plan to submit this, but someone suggested it.)
Caffeine affects the immune system via at least two opposing mechanisms.
Mechanism 1: A2A receptor antagonism (immunostimulatory) Tumors and damaged tissues release adenosine, which engages the A2A receptor on immune cells and signals them to stand down. Caffeine antagonizes (i.e., blocks) this receptor.
Mechanism 2: Raising intracellular cAMP (immunosuppressive) Caffeine also inhibits phosphodiesterase, the enzyme that hydrolyzes (i.e., breaks down) cAMP. cAMP accumulates inside immune cells, which acts as a "calm down" signal.
Note: both mechanisms are dose-dependent. At dietary caffeine levels, A2A antagonism likely dominates, whereas PDE inhibition is weak and mainly relevant at higher concentrations. However, the net immune effect in the tumor microenvironment remains unproven.
---
If you would like to learn more, I can outline a framework for technical folks to ease in and become more informed on cancer. Gaps abound. The more people who understand cancer, the faster we get to cures. Moreover, personalized cancer treatment is the obvious future. Knowledge acquired now may pay off later (but hopefully not needed).
Generally, you rely on experts.
Who typically became experts by adhering to the conventional wisdom set by gatekeepers.
"Science advances one funeral at a time" feels apt.
Sadly, the problem isn't confined to Alzheimer's.
Whenever only a few people decide what is "right," the same pattern of stifled innovation will generally manifest itself not by design or from malice, but because it's hard for a small group to be 100% right on what works and what doesn't -- especially on matters as inscrutable as neuroimmune diseases.
When a topic only has a limited number of experts, those experts become gatekeepers.
Those gatekeepers directly or indirectly control research funding.
Gatekeepers necessarily harbor biases, some right and some wrong, about how the field should progress.
For Alzheimer's, some gatekeepers were conflicted and potentially directed the field in the wrong direction. Only time will reveal AB42's true role.
It's easy to find fault in Alzheimer's.
It's harder to see the general solution to the gatekeeper problem, i.e., how to allocate resources in areas with limited experts.
It’s easy to dunk on VCs, but the herd effect is rational after considering the typical VC’s background, the intense competition for good deals, and the job requirements — to prudently deploy capital.
Who wants to pitch their boss on investing $1-10M in a product no one uses, built by a team of anons?
This is not to defend the process, but merely explain it. It’s not so different from customer marketing. To win a VC, first understand the VC.
Once hired, VCs cannot easily get fired yet they exert immense strategic control.
Nonetheless, many founders interview summer interns harder than VCs.
Heuristic: after removing capital, would you hire the VC to be your boss?
Great VCs are worth the equity and will turbocharge startups. When you find one, don't haggle. Get a fair deal, and get right back to coding.
Bad VCs will destroy companies the same way soccer stars would destroy basketball teams if made the head coach.
For instance, it is not uncommon for cancer studies to design assays around non-oncogenic strains, or for assays to use primer sequences with binding sites mismatched to a large number of NCBI GenBank genomes.
Another example: studies relying on The Cancer Genome Atlas (TCGA), which is a rich database for cancer investigations. However, the TCGA made a deliberate tradeoff to standardize quantification of eukaryotic coding transcripts but at the cost of excluding non-poly(A) transcripts like EBER1/2 and other viral non-coding RNAs -- thus potentially understating viral presence.
Enjoy the rabbit hole. :)
This topic is important and the study interesting, but the methods exhibit the same generalizability bias as the famous Dunning-Kruger study.
The referenced MBA students -- and by extension, the elites -- only reflect 271 students across two years, all from the same university.
By analyzing biased samples, we risk misguided discourse on a sensitive subject.
@dang
Some nuance:
What happens when the methods are outdated/biased? We highlight a potential case in breast cancer in one of our papers.
Worse, who decides?
To reiterate, this isn’t to discourage the idea. The idea is good and should be considered, but doesn’t escape (yet) the core issue of when something becomes a “fact.”
That said, your comment has an implication: in which fields can we trust data if incentives are poor?
For instance, many Alzheimer's papers were undermined after journalists unmasked foundational research as academic fraud. Which conclusions are reliable and which are questionable? Who should decide? Can we design model architectures and training to grapple with this messy reality?
These are hard questions.
ML/AI should help shield future generations of scientists from poor incentives by maximizing experimental transparency and reproducibility.
Apt quote from Supreme Court Justice Louis Brandeis: "Sunlight is the best disinfectant."
Every fact is born an opinion.
This challenge exists in most, if not all, spheres of life.
(To clarify: this is not the fault of scientists. This is a byproduct of a severely broken system with the wrong incentives, which encourages publication of papers and not discovery of truth. Hug cancer researchers. They have accomplished an incredible amount while being handcuffed and tasked with decoding the most complex operating system ever designed.)
1. Nucleotides are a form of tokenization and encode bias. They're not as raw as people assume. For example, classic FASTA treats modified and canonical C as identical. Differences may alter gene expression -- akin to "polish" vs. "Polish".
2. Sickle-cell anemia and other diseases are linked to nucleotide differences. These single nucleotide polymorphisms (SNPs) mean hard attention for DNA matters and single-base resolution is non-negotiable for certain healthcare applications. Latent models have thrived in text-to-image and language, but researchers cannot blindly carry these assumptions into healthcare.
There are so many open questions in biomedical AI. In our experience, confronting them has prompted (pun intended) better inductive biases when designing other types of models.
We need way more people thinking about biomedical AI.
Even Google DeepMind's relabeled MedQA dataset, created for MedGemini in 2024, has flaws.
Many healthcare datasets/benchmarks contain dirty data because accuracy incentives are absent and few annotators are qualified.
We had to pay Stanford MDs to annotate 900 new questions to evaluate frontier models and will release these as open source on Hugging Face for anyone to use. They cover VQA and specialties like neurology, pediatrics, and psychiatry.
If labs want early access, please reach out. (Info in profile.) We are finalizing the dataset format.
Unlike general LLMs, where noise is tolerable and sometimes even desirable, training on incorrect/outdated information may cause clinical errors, misfolded proteins, or drugs with off-target effects.
Complicating matters, shifting medical facts may invalidate training data and model knowledge. What was true last year may be false today. For instance, in April 2024 the U.S. Preventive Services Task Force reversed its longstanding advice and now urges biennial mammograms starting at age 40 -- down from the previous benchmark of 50 -- for average-risk women, citing rising breast-cancer incidence in younger patients.
They are not only unsupported by recent research trends and general patterns in ML and computing, but also by emerging developments in China, which the post even mentions.
Nonetheless, the post is thoughtful and helpful for calibrating investor sentiment.
For ML, start with MedGemma. It's a great family. 4B is tiny and easy to experiment with. Pick an area and try finetuning.
Note the new image encoder, MedSigLIP, which leverages another cool Google model, SigLIP. It's unclear if MedSigLIP is the right approach (open question!), but it's innovative and worth studying for newcomers. Follow Lucas Beyer, SigLIP's senior author and now at Meta. He'll drop tons of computer vision knowledge (and entertaining takes).
For bio, read 10 papers in a domain of passion (e.g., lung cancer). If you (or AI) can't find one biased/outdated assumption or method, I'll gift a $20 Starbucks gift card. (Ping on Twitter.) This matters because data is downstream of study design, and of course models are downstream of data.
Starbucks offer open to up to three people.
Feynman's quote -- "A scientist is never certain" -- is apt for biomedical ML.
Context: imagine the human body as the most devilish operating system ever: 10b+ lines of code (more than merely genomics), tight coupling everywhere, zero comments. Oh, and one faulty line may cause death.
Are you more interested in data, ML, or biology (e.g., predicting cancerous mutations or drug toxicology)?
Biomedical data underlies everything and may be the easiest starting point because it's so bad/limited.
We had to pay Stanford doctors to annotate QA questions because existing datasets were so unreliable. (MCQ dataset partially released, full release coming).
For ML, MedGemma from Google DeepMind is open and at the frontier.
Biology mostly requires publishing, but still there are ways to help.
After sharing preferences, I can offer a more targeted path.
For instance, genomic data that may seem identical may not actually be identical. In classic biological representations (FASTA), canonical cytosine and methylated cytosine are both collapsed into the letter "C" even though differences may spur differential gene expression.
What's the optimal tokenization algorithm and architecture for genomic models? How about protein binding prediction? Unclear!
There are so many open questions in biomedical ML.
The openness-impact ratio is arguably as high in biomedicine as anywhere else: if you help answer some of these questions, you could save lives.
Hopefully, awesome frameworks like this lower barriers and attract more people.
Cancers with established viral etiology or strong association with viruses include:
- Cervical cancer - Burkitt lymphoma - Hodgkin lymphoma - Gastric carcinoma - Kaposi’s sarcoma - Nasopharyngeal carcinoma (NPC) - NK/T-cell lymphomas - Head and neck squamous cell carcinoma (HNSCC) - Hepatocellular carcinoma (HCC)