#346: As Uncertainty Persists In The Crypto Industry, All Eyes Are On Binance, & More

Crypto exchange Binance fell victim to the lens of public scrutiny last week in the wake of FTX’s demise. The collapse of major crypto hedge funds, exchanges, and lenders—including Three Arrows Capital, Celsius, BlockFi, and FTX—has heightened distrust in the ecosystem and galvanized efforts to find a cryptographically verifiable proof of solvency. Proof-of-reserves (PoR) appears to be the new industry standard.

Following that trend, recently Binance emphasized its own proof-of-reserves but has not been able to quell the fear and skepticism of crypto experts and auditors at large. The largest BTC weekly net outflows from Binance in history occurred last week: 71,200 bitcoins in the 7 days ended December 16th.

Binance’s outflows appear to differ substantially from those associated with the unwinding of FTX. FTX bitcoin reserves dropped 87% in 7 days during its collapse, while Binance has seen a relatively modest 10% drop in reserves and, importantly, has honored those withdrawals, as shown below. ARK analysts will continue to monitor this metric as a sign of potential stress at Binance.[1]

Although Binance’s outflows appear to be subsiding, the market remains hypervigilant about its solvency and financial robustness. Amid its outflows, uncertainty around Binance intensified last week after accounting firm Mazars Group (Mazars), a South African auditor responsible for certifying Binance’s proof-of-reserves, withdrew its audit attestation for all crypto companies, including Binance, Crypto.com, and KuCoin. Is Mazars merely seeking to minimize its general crypto reputational risk or does this flag real issues in the cryptoasset space?

Binance appears to be facing its biggest test yet. If it can stand strong amid scrutiny, the exchange would send a signal to the crypto market that it is solvent and backed fully.

Regardless of Binance’s outcome, we believe that bitcoin’s value proposition rests on transparency and decentralization. Those features should withstand all other storms, reinforcing crypto’s fundamental principles of self-ownership, decentralization, and transparency.

[1] We also are monitoring its exchange token BNB and its native stablecoin BUSD, both of which currently appear to be stable.

Base editing is an innovative form of gene editing that makes precise changes to genetic code and corrects mutations that cause diseases like cancer. A base editor converts one DNA letter—or base—into another with a deaminase, an enzyme that performs chemistry on a DNA base. Base editors find the target DNA sequence using a programmable protein—like a zinc finger, a TALE, or a disabled CRISPR-Cas9—that binds to DNA and cannot cut the DNA double helix. Base editors nick only one strand of DNA, not both, creating a different cell response. For perspective, human cells experience thousands of DNA nicks naturally every day, but rarely double-strand DNA cuts. Base editing can disable or enable a specific genetic function by correcting a single “letter” in the genome including mutated DNA letters that cause thousands of genetic diseases.

Base editing seems to have transformed the life of a 13-year-old girl in the UK named Alyssa. Alyssa had been diagnosed with T-Cell Acute Lymphoblastic Leukemia, a type of cancer that affects the immune system. Having bucked multiple rounds of chemotherapy and a bone marrow transplant, Alyssa’s cancer seemed incurable until researchers at the Ormond Street Hospital in the UK used a base editor, called a cytosine base editor (CBE), to make three precise simultaneous edits to a donor’s T-cells, reprogramming them to attack her cancer cells but not her healthy cells.

The CAR-T[1] cell therapy wiped out Alyssa’s cancer: within one month she was in remission and, within six months, cancer-free. While still in early stages of research and development, base editing and multiplexing might provide better CAR-Ts to treat patients with heretofore incurable cancers.

[1] CAR-T Cell Therapy is a type of immunotherapy that uses a patient’s or donor’s T-cells and modifies them in the laboratory so that they recognize and kill cancer cells.

A partnership between MosaicML and the Stanford Center for Research on Foundation Models (CRFM) recently demonstrated the capabilities of industry-specific Large Language Models (LLMs) in biomedicine. Using MosaicML’s Cloud platform, CRFM trained a 2.7 billion parameter LLM on PubMed biomedical data and achieved positive results on the US Medical Licensing Exam’s (USMLE) question-and-answer test. Their research shows that standard LLMs trained on domain-specific data can outperform general-purpose models and compete effectively with expert-designed, domain-specific model architectures.

Called PubMed GPT, CRFM’s model is based on a HuggingFace GPT model with 2.7 billion parameters. For ~$38,000, with a cluster of 128 NVIDIA A100-40GB GPUs, researchers used MosaicML’s Cloud Composer and Streaming Dataset libraries—relatively small datasets—to train PubMed GPT over ~6.25 days for a compute duration involving 300 billion tokens. They processed multiple passes over 50 billion tokens, illustrating that LLMs can be trained quickly and cost-effectively with limited data.

Overall, the results of this partnership reveal the potential for custom LLMs as turn-key solutions for any organization with domain-specific data. Still in research and development, the model does highlight the promise of domain-specific language generation models in real-world applications, but our intelligence suggests that domain-specific models will flourish as companies realize the potential of their proprietary datasets and build products with AI capabilities. More information on domain-specific model training can be found at: https://www.mosaicml.com/.