three things happening in AI that nobody's connecting:1. inference costs dropped 1000x in 2 years. not optimization. commoditization. the margin was gone before most companies had a business model.2. RLHF rewards approval, not accuracy. humans rate confident answers higher. so models learned to be confidently wrong in a way that feels right. sycophancy isnt a bug. its the objective function working exactly as designed.3. MoE routes each token to 2-4 specialized sub-networks instead of firing the
0:17
three things happening in AI that nobody's connecting:1. inference costs dropped 1000x in 2 years. not optimization. commoditization. the margin was gone before most companies had a business model.2. RLHF rewards approval, not accuracy. humans rate confident answers higher. so models learned to be confidently wrong in a way that feels right. sycophancy isnt a bug. its the objective function working exactly as designed.3. MoE routes each token to 2-4 specialized sub-networks instead of firing the
5.8M views2 weeks ago
x.comGopal
I have a question. How is it possible that AFTER Election Day, all ballots being dropped heavily go to the 3rd place candidate, and #1 or #2?What are the statistical odds? Has it ever happened?Seems impossible to me. How about you?
1:11
I have a question. How is it possible that AFTER Election Day, all ballots being dropped heavily go to the 3rd place candidate, and #1 or #2?What are the statistical odds? Has it ever happened?Seems impossible to me. How about you?
793.4K views3 weeks ago
x.comBuzz Patterson
Cheap inference will be most of the compute in future.
0:55
Cheap inference will be most of the compute in future.
27.6K views2 weeks ago
x.comThe AI Investor
Bridging the span between hard determinism and expensive inference spend is whole game. Doing it for the world of bits will be one starting point, and for world of atoms will be another. These two directions will converge and their ardent union will change the world forever.
0:57
Bridging the span between hard determinism and expensive inference spend is whole game. Doing it for the world of bits will be one starting point, and for world of atoms will be another. These two directions will converge and their ardent union will change the world forever.
252.9K views2 weeks ago
x.comib
April 8.Two of three macro signals flipped green. The composite crossed into risk-on.A week later $SPX was back above 7,000.By late April $SPX had run the fastest oversold-to-overbought reversal in over four decades. 🧵
0:11
April 8.Two of three macro signals flipped green. The composite crossed into risk-on.A week later $SPX was back above 7,000.By late April $SPX had run the fastest oversold-to-overbought reversal in over four decades. 🧵
16.7K views4 weeks ago
x.comDeep Inference
$SPY I would like to note... no new high was made today.And a clear RSI divergence and rising wedge has formed & broken to the downside.Make of it what you will...
0:18
$SPY I would like to note... no new high was made today.And a clear RSI divergence and rising wedge has formed & broken to the downside.Make of it what you will...
79.8K views3 weeks ago
x.comDeep Inference
Today we release Rhaister, an elegant statistical model that predicts drug phenotypes in new contexts w/ accuracies comparable to experimental assays.And dropping Emerald Bay, a 2M cell dataset measuring long time-course phenotypes across 1000s of drug-cell line interactions.
0:10
Today we release Rhaister, an elegant statistical model that predicts drug phenotypes in new contexts w/ accuracies comparable to experimental assays.And dropping Emerald Bay, a 2M cell dataset measuring long time-course phenotypes across 1000s of drug-cell line interactions.
57.8K views3 weeks ago
x.comNima Alidoust
1:34
Interesting to hear OpenAI Head of Compute Sachin Katti @sk7037 explain why inference workloads will still be served from giant data centers in Texas rather than smaller edge locations.. The inherent scale benefits of large DCs + lack of available labor makes building smaller DCs closer to major metros inefficient and uneconomic."There are still significant benefits to scale on building this compute. Building 50MW of compute is far more expensive per MW than building a GW of compute at one locat
43.2K views3 weeks ago
x.comRittenhouse Research
0:33
both OpenAI and Google DeepMind are pointing to the same next big AI opportunity: INFERENCE.the bottleneck isn’t just “who can train smarter models” anymore, but also making inference cheap and abundant at scale.
7.1K views2 weeks ago
x.comℏεsam
0:12
"I expect a short term top in June. Then 1-2 months of chop, bottoming late July into early August.""If we can see now what is likely coming within 30-60 days from now, we will be best prepared for when the weather turns to hail." Deep Inference, circa mid to late May
16.6K views3 weeks ago
x.comDeep Inference
1:42
Jensen Huang says every chip startup that bet on cheap inference is dead.For years the AI industry consensus was unanimous.- Inference was easy.- Training was hard.Inference would commoditize.Tiny chips.Cheap tokens.NVIDIA's GPUs would lose the inference market the moment AI shipped.Jensen Huang says the entire prediction was WRONG and illogical."That was always illogical to me," Huang says. "Because inference is thinking. And I think thinking is hard. Thinking is way harder than reading."The mi
7.6K views1 month ago
x.comAI Evolution
0:19
reward guidance is one of the most useful tricks we have for steering diffusions and flows at inference time. you take a learned generative process and nudge it toward high-reward samples, and in principle you're sampling from the reward-tilted measure.but are we really tilting?🤔in practice, people use approximations, and the best one we have is a finite-particle plug-in estimate of the value function. it turns out that this approximation doesn't sample the tilted measure at all.we show that th
7.1K views3 weeks ago
x.comNicholas Boffi
1:39
I just proved that inference futures are no longer a theoretical whitepaper. They are working right now on @solanaWatch the breakdown: I pull a live multi-collateral deposit (USDC + AUTO) from a standard Solana wallet, process it into a dollar-pegged internal clearing ledger, and execute a leveraged 5x long position on `cDEEPSEEK-SEP26` capacity on @autonai_sol I built the CME for the machine economy. @mert @toly @solana
5.7K views1 week ago
x.commankind
0:39
Someone just taught AI to teach itself on impossible problemsLanguage models improve by sampling many answers and keeping the winners On hard reasoning problems,it quietly breaks down. Two flaws cause it. 1. Feedback only arrives once a full attempt passes or fails2. Every candidate stays trapped inside the model's own probability massCorrect answers usually live outside that narrow shell A new paper called BES attacks both at onceIt runs two searches that feed each other: > Forward recombines p
7.6K views1 month ago
x.comAaliya
14:12
From Autoregression to Diffusion: Implications for AI Inference Economics.The central conclusion is that diffusion LLMs are now credible but not yet thesis-sufficient as wholesale replacements for autoregressive LLMs. The base-case architectural outcome is coexistence and hybridization: AR remains the default for frontier general reasoning, open-ended chat, long-horizon tool use, long variable-length prose, and mature enterprise post-training stacks, while diffusion and block-diffusion models wi
1.5K views2 weeks ago
x.comTheValueist
4:10
#SAInflation || South Africa's inflation rises to 4,5% in May 2026. The inflation surge was largely driven by increases in fuel prices. The fuel index recorded a second large monthly increase, leaping by 14,3% to reach an annual rise of 28,7%. Over the past 12 months, prices for petrol increased by 24,8% and diesel by 53,8%.Read more here: https://t.co/k1abkEXJMT#StatsSA #KnowYourStatsZA #CPI #GovZAUpdates
3.7K views2 weeks ago
x.comStatistics South Africa (Stats SA)
0:14
Everything I just walked you through, the bottom, the regime, the liquidity, the strength board, was a fraction of what runs inside.While the furus sell you one call at a time, Aion Analytics hands you the whole engine.Trial is live. Link in bio. Stop guessing.Not financial advice. Mathematically driven models.
679 views1 week ago
x.comDeep Inference
19:01
Research Doesn't Have to Be Hard! Experimental Designs Made Simple (Pt. 1) NCE, CPCE, NCMHCE & EPPP!
27 views2 weeks ago
YouTubeGeneva Tutoring NCMHCE CPCE NCE (PhD S…
See more
Static thumbnail place holder