Context, tokens, KV stores & storage, Solidigm presents at #AIFD8
Solidigm’s presentation at AIFD8 highlighted prompt processing activity and how this relates to context token growth and KV store activities. As KV store grows, offloading parts of it to DRAM, local or networked storage can speed up LLM processing because it doesn’t have to recalculate its KV store every time it stops and waits for…
Keep readingLife hacks: living (phone) life in B&W
Changing your phone display to B&W can reduce your screen time. So what else could use the same hack…
Keep readingWhat R/R tracks can tell us about AI deployment
How R/R track deployment 1850-1950 can explain AI rollout today.
Keep reading
Hammerspace and the Open Flash Platform at #AIIFD3
At AIIFD3, Hammerspace talked about the Open Flash Platform (OFP) initiative, which redefines NAS, for increasing NAND capacities, ubiquitous compute and extreme networking bandwidth, targeted for AI. The DPU-based NFSv3 server will supply up to 24PB in 1RU with coming E2(E) SSDs, optimizing power, capacity and networking to address AI storage performance.
Keep reading
The curse of Scale & AGI
Scale has worked wonders for current foundation models regardless of the cost. But at some point scale can’t help anymore. The time is near to consider alternatives to scale, especially as we attempt to create AGI. We discuss one we think might help.
Keep reading
MinIO Strong Consistency and object gateways #CFD23
During MinIO’s Cloud Field Day 23 presentation they questioned whether object storage gateways offered strong consistency for large objects. The post talks about my views on this.
Keep readingAGI, SuperIntelligence and “The Last Man”
Nietzsche’s, the “Last Man”, is a future where automation & technology leads to a life of unchallenged leisure, devoid of struggle and growth. AGI and SuperIntelligence, at best will lead to the Last Man, at worst, to the end of humanity. And we are closing fast on one of these two possibilities.
Keep reading
AlphaEvolve, DeepMind’s latest intelligence pipeline
Deepmind’s AlphaEvolve uses genetic and other techniques via GeminiFlash and GeminiPro to evolve code for the better. It evaluates changes using supplied scoring algorithms and has resulted in many advances in mathematics, coding, digital design, etc.. It can be used to improve any task that can be described in code which can be evaluated for…
Keep reading
Reward is all you need – part 2, AGI part 12, ASI part 3
We discuss a new article, “Welcome to the Era of Experience” which depicts an approach to attain super-human intelligence for LLM agents.
Keep readingSomething went wrong. Please refresh the page and/or try again.