🚨 GPT-5.6 features flag can be enabled in the codex > So we can see ui but can't use the model directly > Only trusted users are given access for now Personal life update: my adhd get worse with vacation as it fuck up my routine that's why I was not active lately, suggest something to solve this 🫂

Everyone's chasing bigger GPU clusters, but the real bottleneck in most AI training setups isn't compute — it's storage. What caught my attention about @KAYTUS_ 's new All-QLC flash architecture is how directly it addresses this. When you're running 10,000 GPUs in parallel, feeding them data fast enough becomes the actual engineering challenge. Traditional storage just wasn't built for this scale. The QLC approach makes a lot of sense here. AI training workloads are overwhelmingly read-heavy — loading datasets, checkpointing models, shuffling batches. Paying a premium for TLC write endurance doesn't really add up when your workload barely writes. QLC delivers more capacity per dollar with better economics at scale. The specs back it up: 10 TB/s aggregate bandwidth, 100 million random-read IOPS, and GPU utilization above 95%. That last number matters most — idle GPUs are expensive GPUs, and even a small utilization gain across 10,000 GPUs translates to massive savings. The 70% lower 5-year TCO compared to TLC is also worth paying attention to as clusters continue scaling. The AI infrastructure race is evolving from "who has the most GPUs" to "who keeps them productive." Storage is one of the most overlooked pieces of that puzzle, and KAYTUS is making a compelling case for solving it.

Fable 5 fails this test , first test that it fails > need to work on vision > google models can easily solve this @theo this is example that you want fable failure on prompt , i think vision is still very bad compared to gemini correct answer - No , there is tick on muffin

90% of legal AI is just “summarize this PDF” Ivo Benchmarks is different it checks the clause in front of you against your company’s actual negotiation history > how often you accepted it > what language you used before > whether to accept or push back this is the real vertical AI unlock

Bloome is actually smart for coding workflows i added Claude Code, Codex, DeepSeek and a Project Lead agent into one group chat each agent independently built their own version one challenged the logic, one defended, one fixed missing details and the final website stayed inside the same thread no copy paste between agents no lost context just one group chat for getting work done @Bloome_im

🚨 Sonnet 5 < get opus 4.7 tokenizer > , but the hidden thing is tokenizer change same text can map to 1.0x–1.35x more tokens , so more prices so Anthropic made intro pricing “cost neutral” lmao 😂

🚨 Sonnet 5 it's been a long time for a sonnet update If rumours are right then it will be released today ( this time I have no idea neither I have tested it yet as of writing this tweet ) Once I will get my hand I will start uploading tests

🚨 Exclusive: new Claude app strings mention Fable 5 credits + identity verification Fable 5 looks like it’s being put behind usage credits, billed outside your plan. one string says: “Your credits will be added once your identity is verified.” the weird part? Anthropic said ID verification wasn’t a Fable thing and was only for flagged accounts. but these strings showed up with the Fable 5 credit changes

I am waiting for GPT -5.6 Sol to be used as default model in daybreak I have two questions > Do we get cyber edition or normal Sol > Can we have less restriction in trusted access compared to 5.5 cyber

GLM 5.2 is first chinese model that made 3d watch without any major mistakes 🤯 @Zai_org you guys cooked very hard with this model , now only thing that we need is the big fat model , tell us the time when we can expect This is one of the hardest thing to one shot

Everyone's chasing bigger GPU clusters, but the real bottleneck in most AI training setups isn't compute — it's storage. What caught my attention about @KAYTUS_ 's new All-QLC flash architecture is how directly it addresses this. When you're running 10,000 GPUs in parallel, feeding them data fast enough becomes the actual engineering challenge. Traditional storage just wasn't built for this scale. The QLC approach makes a lot of sense here. AI training workloads are overwhelmingly read-heavy — loading datasets, checkpointing models, shuffling batches. Paying a premium for TLC write endurance doesn't really add up when your workload barely writes. QLC delivers more capacity per dollar with better economics at scale. The specs back it up: 10 TB/s aggregate bandwidth, 100 million random-read IOPS, and GPU utilization above 95%. That last number matters most — idle GPUs are expensive GPUs, and even a small utilization gain across 10,000 GPUs translates to massive savings. The 70% lower 5-year TCO compared to TLC is also worth paying attention to as clusters continue scaling. The AI infrastructure race is evolving from "who has the most GPUs" to "who keeps them productive." Storage is one of the most overlooked pieces of that puzzle, and KAYTUS is making a compelling case for solving it.

Google This is what I love it , now my notebookLM experience will be top notch New skills unlock every day I am very happy this feels so personal , this is what I needed in my childhood, even today this is the best tech for me 😭 Gemini Omni made video to explain photosynthesis Go follow @mrfanduuuuu

Holllllyyyyyyyy @GeminiApp cooked 😳😳 🚨 Gemini Omni: New video model Here is the first output and see the text coherence , if this is not nano banana moment of video then what is ?? direct link for those who believes otherwise in comments

MiniMax M2.7 multi agents system made a water physics with proper temperature control website i have shared the codepen link and also @MiniMax_AI now please cook big model 1T or 2T , i have high hope from you guys