The transformer technology used in today’s large language models has dramatically changed what AI can do. Indeed, these models process long pieces of text all at once, which helps them understand context quickly. However, transformers struggle with tasks that require step‑by‑step reasoning or tracking details over time. To fix this, researchers have started pairing transformers with older, more sequential approaches that remember information as they go, similar to how humans track details in a conversation. Interestingly, tech giant IBM (IBM) recently revealed one of the most promising solutions.
TipRanks Cyber Monday Sale
- Claim 60% off TipRanks Premium for data-backed insights and research tools you need to invest with confidence.
- Subscribe to TipRanks' Smart Investor Picks and see our data in action through our high-performing model portfolio - now also 60% off
In fact, a new model, known as PD‑SSM, restructures how state space models (SSMs) store information, thereby allowing them to track sequences accurately without losing speed. In testing, PD‑SSM performed far better than existing models on tasks that require remembering and updating information over time, which include predicting what comes next in real‑world time‑ordered events.
It’s also worth noting that these findings bring back debates that go back to the 1950s, when linguist Noam Chomsky described how some problems require step‑by‑step “state tracking.” In addition, basic tests, such as checking whether a string of ones and zeros adds up to an odd or even number, are used to expose weaknesses in AI systems, and transformers often fail when the sequence gets long enough. IBM’s model not only succeeded on these tough tests but also worked well on practical tasks like analyzing heartbeats. As a result, IBM plans to bring PD‑SSM technology into its Granite‑4 models.
Is IBM a Buy, Sell, or Hold?
Turning to Wall Street, analysts have a Moderate Buy consensus rating on IBM stock based on eight Buys, five Holds, and one Sell assigned in the past three months, as indicated by the graphic below. Furthermore, the average IBM price target of $300.58 per share implies that shares are trading near fair value.
