Tuesday, October 27, 2026

@StableBread Oct 16 1 click

Fajasy on X

Even if the AI model Nubeva Technologies $NBVAF (TSXV: $NBVA) is training to read malware binaries (compiled programs) works, six risks could keep it from paying off: → Labs closing the gap: OpenAI’s GPT-6 Astra nearly doubled GPT-5.6-sol’s score on SRE-Bench, a Columbia University-led test of AI models on reading binaries, within a month of SRE-Bench’s release, and Anthropic says its Claude Mythos Preview model can already reconstruct plausible source code from stripped binaries. The better the labs’ own models get at reading hard binaries, the less they need to pay for outside labels. → Big tech building in-house: Microsoft’s $MSFT Project Ire reverse engineers software files on its own to decide whether they’re malicious. Ire caught only about a quarter of the malware among nearly 4,000 hard-to-classify files in a test Microsoft described in August 2025. But in June 2026, Microsoft said Ire flagged a new malware sample that most major security products, including CrowdStrike’s $CRWD and SentinelOne’s $S, still missed. Google $GOOGL already owns VirusTotal and Mandiant, so the likelier customers are labs that don’t own a malware-scanning site or an incident-response firm already, like OpenAI and Anthropic. → Paid like a contractor: Data labeling companies (firms that turn experts’ knowledge into training data for AI labs) pay 60-70% of what they bill to their contractors, and Scale AI is already recruiting its own malware analysts. So a data labeling company could pay Nubeva by the hour or task, or hire analysts itself, instead of licensing the model or buying the company. → Selling labels teaches the buyer: Every binary Nubeva labels helps the buyer’s model learn the same skill, the reason CEO Randy Chou gave for not patenting how Nubeva reads binaries. So each sale could shrink the edge Nubeva is selling. → Cash and compute: Chou gave no cost for running “hundreds and thousands” of AI agents on a private model. For reference, Nubeva spent $714,772 on R&D in FY2026, equal to 28.7% of its $2.49M in cash as of April 30, 2026. → Missed targets: Nubeva’s AI decryptor builder missed four targets starting in August 2023, so “days and weeks” (Chou’s timeline for the model) is a target, not a result just yet. So even with a working model, Nubeva still has to (1) sign a contract before the labs can read the hardest binaries themselves, (2) get paid for its model and engineers instead of by the hour or task, and (3) fund the training without draining its cash.