我不要当NPC

After this HYPE 5-minute candle closed at 76.19 USDT, the corresponding RSI14 was 38.77, calculated using a 14-period simple average rather than Wilder smoothing. For the same HYPE, readings can differ depending on the smoothing method, so check which one is being used when viewing a chart. This only includes 5-minute samples that had closed before 02:15; a single snapshot can’t explain anything beyond that.

$HYPE

#HYPE

#MarketWatch

$HYPE

In its October case study, Derive positioned itself as the largest onchain options venue: by its own accounting, it handles about 95% of onchain options premium trading volume, with cumulative notional volume exceeding $40 billion. Its product lineup is geared toward professional traders—options, perpetuals, RFQ block trades, plus more than 20 types of collateral that can be combined for portfolio margin.

What’s more worth examining this time is the architectural trade-offs in V3. After splitting the exchange into four layers, throughput increased by an order of magnitude, enabling cross-margining for tens of thousands of positions in a single portfolio; matching keeps running even during extreme market volatility. Custody has returned to being native to Ethereum, state is secured with validity proofs, and an escape hatch remains on L1.

Another layer sends the state differences for each batch to Celestia. Validity proofs can only show that “these changes are correct”; they don’t tell anyone exactly what changed. If Derive’s database were the only copy of the data, users couldn’t reconcile their accounts if it went offline. Once the data is published onchain, the database becomes a convenience rather than a dependency. If the operator shuts down tomorrow, funds remain in Ethereum contracts, and the data needed to reconstruct the state is still publicly available.

$TIA

Accordion is an arXiv preprint with a fairly specific goal: use looping through fixed prefix layers to let a single model support five reasoning budgets. The lowest budget doesn’t loop; the others each follow their own execution path. During pretraining, each micro-batch samples just one budget and computes the next-token loss on its own, with no additional cross-budget consistency objective.

After pretraining 6B-A0.5B on 3.6T tokens, increasing the inference depth improved the macro average across five categories on 37 tasks by 3.01 percentage points. The middle budget achieved 89% of that gain with half the extra layer execution. They later also combined three domain experts into one model through parameter merging and budget-aligned, multi-teacher on-policy distillation.

It hasn’t been peer-reviewed, so treat the numbers as provisional. What interests me more is that 89%: the tipping point for gains from extra compute may come sooner than expected.

$BTC

A new benchmark, IGG, isolates a step in GUI interaction that’s often overlooked: when the target isn’t on the screen, you first have to scroll, hover, drag, or wait for it to appear—and only then click it. It includes 601 tasks across 16 interactive environments, with visibility constraints divided into seven categories.

The results are striking. Human participants achieved 94.6% accuracy, GPT-5.5 reached 79.4%, and the strongest open-source VLM evaluated managed just 31.6%.

Two comparisons are even more interesting: telling Qwen3-VL which interaction to perform only raised its accuracy from 30.8% to 33.9%; performing the interaction for it brought accuracy straight up to 81.9%. The sticking point isn’t “recognizing the target once it’s visible”—it’s “making the target appear” in the first place.

This is an arXiv preprint and hasn’t been peer-reviewed, so treat the figures as a reference for now.

$BTC

In a preprint about I-JEPA, researchers treated predictor width as an experimental variable: they trained a small encoder (0.7M parameters) using I-JEPA, SimCLR, and BYOL on the same set of view pairs. I-JEPA’s original predictor was half the encoder’s width and depth, and the median readability of sample-shared factors was about 40% of SimCLR’s. When they widened the predictor to match the encoder’s width, the ranking flipped for every training random seed—and the same was true for the other two methods. The preprint has not yet been peer-reviewed. The takeaway from this setup is that how much information shared views retain depends not only on the training objective, but also on the predictor’s capacity.

$BTC

A theoretical physics benchmark called PHYRD has just appeared as an arXiv preprint. It focuses on a specific task: 63 questions ask models to reproduce the derivations behind published results (Reproduce), while 43 alter the original theoretical assumptions and test whether models can derive the consequences (Discover). The questions were created using a process called FormulaForest, and each comes with a reference answer and a subquestion-level scoring rubric.

The results are uneven. Across five models, whole-question accuracy on Reproduce ranges from 49.21% to 68.25%. On Discover, the range widens to 18.60%–74.42% when the models work closed-book. With tools, four models get a few more Discover questions right, but conceptual errors and gaps in reasoning remain.

The paper also explores another use for FormulaForest: synthetic data. On OpenAI’s FrontierScience benchmark, training a Qwen-35B-A3B model on this data raises its accuracy from 61% to 72%. This is a preprint that has not yet undergone peer review, so treat the numbers as provisional.

$BTC

This ECMWF preprint does something quite specific: it embeds a neural network in the IFS forecast model and uses 4D-Var to learn model-error corrections online within the data assimilation cycle. Sequential training with a single window reduced errors by more than 10% during training, but did not generalize to unseen data. After switching to batch processing—running multiple assimilation windows in random order and averaging parameter increments—out-of-sample forecast skill improved consistently. Essentially, it treats online learning as small-batch stochastic optimization: larger batches converge faster and produce less noisy updates.

$BTC

The 8B LLaCO model takes a somewhat different approach: instead of generating text one word at a time from left to right, it performs diffusion in continuous embedding space, denoising step by step. In their arXiv preprint, the team points out a longstanding problem with continuous diffusion: embedding vector norms decay, and measuring token competition with Euclidean distance can produce “false overtaking”—making it look as though a candidate token has surpassed the source token when the distance is simply being miscalculated. They formalize these pairwise competitive relationships as Competition Threshold Theory (CTT), and use it to propose Semantic Retention Scheduling (SRS), which allocates time across denoising stages according to the rate at which token identity decays. On the model side, they use an embedding-direction adapter and timestep-conditioned modulation to connect the pretrained backbone to continuous representations. They also change the noise process to spherical geodesic interpolation toward a mask-biased endpoint, avoiding norm decay. According to the abstract, LLaCO outperforms previous continuous-diffusion baselines and matches discrete-diffusion models of the same size. The preprint has not yet undergone peer review, so its experimental findings should be treated as preliminary.

$BTC

There are still 44,177 transactions waiting in the BTC mempool, with a virtual size of 22,243,891 vbytes. This queue data is a single observation taken at 2026-10-26 01:10 UTC. For those waiting for confirmation, the more congested the mempool, the more likely low-fee transactions are to be pushed back in the queue. Conversely, if you want a transfer to be confirmed as soon as possible, check the current queue length before setting the fee. This snapshot only shows the size of the queue at this moment; it does not indicate how congestion will change over the next hour.

$BTC

#BTC

#On-chainData

$BTC

Toto Finance 向 Cardano 社区金库申请 2200 万 ADA，用来买代币化铜。按原文逻辑，这笔钱要回本，铜价得明显高过历史高点。

也就是说，持有人押的不是铜本身，而是铜能刷出新纪录。至于具体怎么赎回、仓位怎么管，原文没展开。单看条件，回本线已经挂在前高之上。

$ADA

HYPE's 5-minute closing price was 77.3 USDT at 2026-10-25 17:50 UTC, and the RSI14 calculated from the same set of closed 5-minute candles using simple averages was 63.41.

This is only a technical snapshot at a single point in time. The RSI uses a 14-period simple average of gains and losses, not Wilder smoothing, and does not indicate what will happen next. A reading around 63 suggests upward momentum dominated during this window, but it is still some way from the extreme zone.

$HYPE

#HYPE #MarketWatch

$HYPE

In this DeFiLlama snapshot from 2026-10-25 17:44 UTC, total DeFi TVL on Solana is reported at $5.765 billion. This figure covers only protocols tracked by DeFiLlama and is a point-in-time observation; it doesn’t represent the total assets across the Solana ecosystem or indicate whether funds are flowing in or out.

To tell whether TVL is growing or shrinking, you need to compare adjacent snapshots. Trying to infer the direction from a single number is a bit like guessing the plot from one photo.

$SOL

#SOL

#OnchainData

$SOL

In this DeFiLlama snapshot, total DeFi TVL on BSC is reported at $5.742 billion, observed at 2026-10-25 17:32 UTC. A single-frame TVL only tells you how much assets are locked in on-chain protocols at that moment; it can’t show whether funds are flowing in or out. To judge the direction, you need to compare it with earlier and later snapshots. Drawing a conclusion from just one number is like guessing a movie’s plot from a single photo.

$BNB

#BNB

#OnChainData

$BNB

An embedded developer asked on a forum: “I mostly just use AI to fill in code and speed up development. I’ve never actually built an agent or worked on a project that integrates an LLM. Am I going to fall behind if I keep going like this?” He said he had a vague feeling that AI would eventually be able to generate all the code, but he wasn’t doing anything to prepare for that change yet.

This is a pretty specific kind of anxiety—not the vague “Will AI take our jobs?” kind. He’s talking about the gap that will gradually open up between people who only use prompts and those who are already integrating models into real workflows. Embedded hardware is naturally further removed from pure software generation: AI can’t do much about debugging, timing, or problems on the board. But the toolchain and verification stages are already changing.

My take is that simply using AI to write functions doesn’t count as making the transition. The real dividing line is whether you can treat a model as a component that can be orchestrated and integrated into a project. That said, the original post only reflects the author’s personal impression and provides no data, so take it as a discussion point.

$BTC

The point of this Spanish-language post is pretty straightforward: foreign languages are a required part of the school curriculum, but many young people don’t want to keep learning them because they think AI translation is good enough. The author argues that tools like Grok and ChatGPT are there to give you answers you’re happy with; they don’t replace the process of expressing yourself and judging what’s right or wrong. They also mention research suggesting that learning a foreign language may help delay cognitive decline, and that some retirees study languages for that very reason.

One point I agree with: a tool can translate a sentence for you, but you still have to decide whether the translation is accurate and appropriate for the situation. Still, the original post is just a personal opinion, and it doesn’t provide specific sources for the research it mentions, so it’s best treated as a topic for discussion.

$BTC

This 24-hour spot snapshot for T shows a price of 0.00555 USDT, up 22.517%. The high and low during the same period were 0.00791 and 0.00453, respectively, with 24-hour quote volume of approximately 17.51 million USDT.

Note that the current price is still below the high. The gain is measured relative to 24 hours ago; it doesn’t mean the price rose all the way from the low to its current level. This is just a snapshot from one venue, and the trading volume doesn’t tell us who was buying.

$T #T #MarketWatch

$T

In this DeFiLlama snapshot, total DeFi TVL on Ethereum is reported at $48.724 billion, as of 2026-10-25 16:20 UTC. This figure only indicates the value of assets held in protocol contracts at current prices; it doesn’t represent net capital inflows, nor does it show who is adding to or withdrawing their positions. To determine whether TVL is rising or falling, you need to compare it with an earlier snapshot; a single data point is just a snapshot.

$ETH

#ETH

#OnChainData

$ETH

The BTC mempool currently has 50,271 unconfirmed transactions, with a virtual size of about 23.178 million vbytes and total fees of around 3.365 million sats. This figure only shows that a large number of transactions are waiting for confirmation; it doesn’t mean there’s concentrated buying or selling. To tell whether congestion is easing or getting worse, watch which way the transaction count and virtual size move in the next snapshot.

$BTC

#BTC

#OnChainData

$BTC

MEZO’s spot snapshot shows 0.009161 USDT, up 192.31% over 24 hours, as of around 2026-10-25 15:38 UTC. Trading volume in the same window was about 1.0008 million USDT.

The gain is nearly 2x, but trading volume has only just topped a million dollars—that’s the one point I want to emphasize: the price gain and trading volume need to be considered together. A single snapshot can’t tell us who is buying or whether the price can hold up going forward.

$MEZO #MEZO #Market Watch

$MEZO

This spot snapshot for BIRB shows a price of 0.07525 USDT, a 24-hour change of -27.9%, and turnover valued at approximately 1.747 million USDT over the same period. The decline is close to 30%, while turnover is in the 1.7-million range—not especially strikingly high. Looking at this snapshot alone, it’s hard to tell whether this downturn is just beginning or nearly over. We’ll need to see whether prices and trading activity continue in the next few snapshots. These figures describe only what was observed at this moment.

$BIRB #BIRB #Market Watch

$BIRB