This week in bio:

1. 100 billion parameter protein language model (the largest ever trained), trained for 6+ months on 96 NVIDIA A100s (likely costing millions of dollars) to achieve SOTA on a range of protein optimization tasks

2. wildDISCO, a new technique for whole-body cell mapping in mouse using off-the-shelf antibody immunolabeling, enabling a new method for building cellular and structural maps in various disease states

3. adopting product management methodologies from tech to the wet lab to improve research productivity

4. an overview of different subtypes of diffusion models in generative chemistry for drug design, and a call for more rigorous evaluations beyond simplified metrics that encourage "SOTA leaderboard" climbing rather than truly moving the needle on developing more drug-like molecules

5. single-photon camera for the the non-invasive measurement of neural activity

This is HUGE. Keras launching Keras Core, a new library that brings the Keras API to JAX and PyTorch in addition to TensorFlow.

It enables you to write cross-framework deep learning components and to benefit from the best that each framework has to offer.

Claude-2, Anthropic's shot at GPT-4, has arrived

It's cheaper than GPT-4 and far stronger in reasoning & coding than its older self.

You can try it out at claude.ai

Things you should know:

1. On standard exams, it's not quite at GPT-4 yet but catching up fast compared to v1.3. Winner in bracket:
GRE verbal: 165 vs 169 (GPT-4 wins)
GRE writing: 5 vs 4 (Claude)
GRE quantitative: 154 vs 163 (GPT-4)
USMLE: ~67 vs ~85 (GPT-4)
Bar: 76.5 vs 75.7 (Claude)

2. On reasoning benchmarks,
HumanEval coding: 71.2% vs 67% (Claude wins). Meanwhile GPT-3.5 only scores 48%. Claude's coding ability has improved dramatically.
GSM-8K grade-school math: 88% vs 92% (GPT-4). Claude improves from 85.2 -> 88 vs its prior version.

3. Claude 2 (100K context) is 4-5x cheaper than GPT-4-32K! Prompt tokens cost $11 vs $60/Million, and completion costs $32 vs $120/M, assuming similar tokenization length.

4. Claude-2's knowledge cutoff is in early 2023, while GPT-4 is Sept. 2021. So it's got much fresher memory.

5. Lobotomy is very aggressive. Claude 2 is 2x better at giving harmless responses than v1.3. Anthropic worked with the Alignment Research Center (ARC) and external red teamers for safety audits.

6. 10% of training data is non-English.

7. Claude's long context means that you can upload entire papers and code files to ask for summary or bug fix.

Big caveat on the standardized exam: the prompting protocols may be very different, and there're no error bars on a large number of exams. The comparison may not be statistically significant.

Today, the European Commission has adopted a new strategy on Web 4.0 and virtual worlds to steer the next technological transition and ensure an open, secure, trustworthy, fair and inclusive digital environment for EU citizens, businesses and public administrations.

"The Web 4.0 and virtual worlds will bring benefits for health, contribute to the green transition and better anticipate natural disasters. But we need to have people at the centre and shape it according to our EU digital rights and principles, to address the risks regarding privacy or disinformation. We want to make sure Web 4.0 becomes an open, secure, trustworthy, fair and inclusive digital environment for all." - Margrethe Vestager, Executive Vice-President for a Europe Fit for the Digital Age.

This is HUGE. Max Hodak the founder of Science Corp introduced current revision of the science eye implant

https://science.xyz/products/vision

The United States announced the first criminal case involving attack on a smart contract operated by decentralized exchange.

AHMED, 34, of New York, New York, is charged with wire fraud and money laundering, each of which carry a maximum sentence of 20 years in prison.

Shakeeb Ahmed, who was a senior security engineer at an international technology company, used his expertise to defraud the exchange in solana and its users and steal approximately $9 million in cryptocurrency. After he stole the fees he never legitimately earned, AHMED had communications with the Crypto Exchange in which he decided to return all of the stolen funds except for $1.5 million if the Crypto Exchange agreed not to refer the attack to law enforcement.

Though the DEX hasn’t been named in the indictment, it may be related to the Solana-based liquidity protocol Crema Finance hack that happened at that time last year, when a thief made away with $9 million in digital assets in a flash loan attack but then returned most of the cash.

Cryptoassets. The BIS released a report entitled ‘The crypto ecosystem: key elements and risks’, which was submitted to the G20 Finance Ministers and Central Bank Governors.

Highlights

This report reviews the key elements of the crypto ecosystem and assesses their structural flaws.

There are three main takeaways.

1. due to underlying economic incentives, the crypto ecosystem is characterised by congestion and high fees, which lead to fragmentation.

2. despite an original ethos of decentralisation, crypto and decentralised finance (DeFi) often feature substantial de-facto centralisation, which introduces various risks.

3. while DeFi mostly replicates services offered by the traditional financial system, it amplifies known risks.

Moreover, as DeFi does not finance activity in the real economy, its growth is driven by the speculative influx of new users, with substantial risks to investors.

The report outlines policy options to mitigate the multiple risks crypto poses to investors, the traditional financial system and the economy at large.

Roblox is coming to Meta QuestVR. Open Beta for developers coming soon on App Lab.

Elon Musk introducing xAI

This is his answer to OpenAI and competitors, built on employees who just left OpenAI, Microsoft, Google, DeepMind, etc.

The iconostasis of scientists is honorary, although somewhat strange. There are also many people on the list who are known in circles where issues of ethics and safety of AI are discussed.

Separately, it is noted that xAI is not associated with X Corp in any way, i.e. twitter. But this is a hoax, because even on the first page, all the links lead to the twitters of the participants, and not to their other pages.

a16z released a massive open-source companion AI repo on GitHub

Chatbot companions are taking off. They have a clear product-market fit.

NASA uses VR and AI to create virtual data worlds

As part of its data visualization research, the US space agency NASA is experimenting with virtual and mixed reality at its Jet Propulsion Laboratory in Southern California.

The aim is to improve the analysis and interpretation of complex data, particularly in relation to the size and distance of geological features on Mars.

Virtualitics' main product is software that combines artificial intelligence with 3D visualization. It allows users to see patterns and relationships in their data faster than is possible with traditional 2D charts.

An excellent narrative by Nandan Nilkeni on India’s Digital Transformation

«India is innovating a different model for growth. A technology-led model that is collaborative, equitable, and democratises opportunity at population scale. As wealth won’t trickle down, so we are raising the floor».

Wallet Pay enables Bitcoin, Tether and Toncoin payments on Telegram

Users will be able to make payments to merchants using Tether, Toncoin and Bitcoin directly within the Telegram app.

Payments can be made via merchants’ bots or web applications on Telegram using Wallet Pay.

NEW Artificially Grown Miniature Brain

Researchers at University of Michigan developed a method to produce artificially grown miniature brains, human brain organoids, free of animal cells.

Engineered EECM matrices represent a major advancement in neural engineering
- they have the potential to significantly enhance the structural, cellular, and functional diversity that can be achieved in advanced brain models.

Key Findings:
- The new brain organoids are grown using an engineered extracellular matrix without animal components, overcoming the previous method’s variability issues.
- The lab-grown organoids using this new method showed enhanced neurogenesis compared to previous studies.
- This method allows for potential reprogramming with cells from patients suffering from neurodegenerative diseases.

Significance:
- This will allow for more accurate replication of human brain conditions and
- will improve the way neurodegenerative conditions are studied
- Will guide personalized treatment of neurodegenerative diseases such as ALS, Alzheimer’s and other neurologic conditions

“This advancement in the development of human brain organoids free of animal components will allow for significant strides in the understanding of neurodevelopmental biology,” said senior author Joerg Lahann, Ph.D., director of the U-M Biointerfaces Institute.

Full summary of the xAI Spaces. Elon and the xAI team have announced:

xAI is aiming to release details of its first product within a few weeks!

- xAI will release tools to the public as soon as possible

- will combine text, image, and real world AI

- Details of the first product release will be out in a few weeks

- Any ordinary person will be able to vote on xAI's future.

Elon: goal of xAI is to build a 'good AGI'.

Best way is to build an AI that is maximally curious and interested in humanity.

Elon: Questions for xAI to answer - the nature of gravity, Fermi paradox (where are the aliens?). To call something AGI it will need to solve atleast one fundamental problem.

Greg Yang adds the AI could give another perspective to existing theories aswell.

Elon: Two metrics to track are the ratio of biological to digital compute globally (eventually biological will be under 1% of the total), and total energy created per person (increasing exponentially).

Regarding xAI as a company, things to note are:

1) It will remain a relatively small team, and with a culture of being able to challenge eachother
2) It will aim to ship product quickly.

ALX: Is xAI in competition with OpenAI or Google's Bard?

Elon: We're in competition.

Elon says the company is still embyronic so there aren't full details of what's to come, but the goal is to make useful AI for consumers and businesses.

How will xAI use Twitter's data?

Elon: Nearly all AI companies have been using Twitter's data - in almost all cases, illegally. There were multiple entities trying to scrape all tweets ever made within a matter of days, which was bringing Twitter's system to its knees.

xAI will use the public tweets for training, text training but also video and image training aswell.

At some point it will run out of human data. xAI will have to generate content and self-assess content - self-generated content will be the path to AGI.

Elon: For xAI, on physics questions the AI will give an objective answerr, but for more subjective topics the AI could give a range of answers depending on what your assumptions are.

Elon: Any cooperation with Tesla will be at arms length since Tesla is a public company, but there will be various links, and xAI would help Tesla's self driving aswell.

Ani Aslanyan: Will xAI have a 3D world knowledge?

Elon says it will be able to understand the physical world - e.g, it could tell you where to find a coffee at an airport.

Isaacson: To what extent will xAI be involved in real world AI?

Elon: xAI can help Tesla and vice versa. xAI can be assisted by Tesla, which has the world's leading real world AI team.

Need to be very careful with Optimus - a hardcoded way of turning the bot off.

This company plans to transplant pig hearts into babies next year

A biotech company called eGenesis is experimenting with transplanting the hearts of young gene-edited pigs into baby baboons as part of a study that could pave the way for similar transplants in human babies.

It hopes to transplant pig hearts into babies with serious heart defects as early as next year, in a bid to buy them more time to wait for a human heart.

The company has developed a technique that uses the gene-editing tool CRISPR to make around 70 edits to a pig’s genome. In theory, these edits should allow the organs to be successfully transplanted into people.

The practice is proving more difficult. The team is planning to test with 12 infant baboons, but of the two surgeries that have been performed so far, neither animal survived beyond a matter of days. Still the company, and others in the field, remain optimistic.

Ethereum-backed stablecoin developer Ethena raises $6M

The project took inspiration from an Arthur Hayes article. He has called for the creation of an algorithmic stablecoin (NakaDollar, or NUSD) backed by bitcoin derivates.

Ethena has developed a stablecoin, USDe, that maintains its dollar peg by using collateral from users to hedge price exposure by shorting ethereum using perpetual swaps.

AI systems sometimes make mistakes. To enable their use in safety critical domains like healthcare, DeepMind and Google AI introduced ‘CoDoC’ - an approach that enables human-AI collaboration by learning when to rely on predictive AI tools, or defer to a clinician.

The South Park Commons Founder Fellowship gives founders the best chance possible.

Applications are open for $1M and the -1 to 0 support that's produced 145 companies worth $35B.

Founder Fellows get:

1. $400k for 7% on a SAFE on acceptance + $600k guaranteed in your next venture round

2. Bespoke support from SPC partners in a small cohort

3. SPC community membership, office space, custom curriculum, personalized network intros & more