Liuseco

Enginerd

[Liuseco @Liu_xoke]

@tabgnuh Tập làm quen đi thầy

[Kimi.ai @Kimi_Moonshot]

Introducing 𝑨𝒕𝒕𝒆𝒏𝒕𝒊𝒐𝒏 𝑹𝒆𝒔𝒊𝒅𝒖𝒂𝒍𝒔: Rethinking depth-wise aggregation.

Residual connections have long relied on fixed, uniform accumulation. Inspired by the duality of time and depth, we introduce Attention Residuals, replacing standard depth-wise recurrence with learned, input-dependent attention over preceding layers.

🔹 Enables networks to selectively retrieve past representations, naturally mitigating dilution and hidden-state growth.
🔹 Introduces Block AttnRes, partitioning layers into compressed blocks to make cross-layer attention practical at scale.
🔹 Serves as an efficient drop-in replacement, demonstrating a 1.25x compute advantage with negligible (<2%) inference latency overhead.
🔹 Validated on the Kimi Linear architecture (48B total, 3B activated parameters), delivering consistent downstream performance gains.

🔗Full report:
https://t.co/u3EHICG05h

[Liuseco @Liu_xoke]

@blvckvux Vietnam foood (rich people)

[Liuseco @Liu_xoke]

@pvergadia Exactly. The irony is that natural language is so fluid and ambiguous that trying to build a 'stable' system with it is like building on sand.

[Liuseco @Liu_xoke]

@zorikgekhman Very cool sharing! As reasoning tokens add VRAM/latency pressure, do you see Adaptive Computation (System 1 vs 2) as the ultimate solution for production LLMs?

[Zorik Gekhman @zorikgekhman]

New paper 🚨

We know that reasoning helps when step-by-step solutions are natural, for example in math, code, and multi-hop factual QA. But why should it help with factual recall, where no complex reasoning steps are needed?

1/🧵

[Liuseco @Liu_xoke]

@blvckvux Chúa sinh ra bitcoin đó, không biết à

[Liuseco @Liu_xoke]

@advaith @AlexCooperWG Why are they bad words bruh

[crashout @0xCRASHOUT]

many of today's protocol tokens are designed primarily for growth speculation and governance. this leads to a token model that is vertically integrated and "app-bounded."

for example, the value of tokens like UNI or AAVE is close tied to the success of their individual protocols, rather than the growth of the broader category they belong to.

PROVE introduces a different paradigm: a "market-bounded" asset.

instead of being tied to a single application, PROVE is positioned to capture economic value of the entire horizontal proofs market. this model is inherently agnostic to which L2, chain, dapp wins, as PROVE benefits from the ever-growing aggregate demand for proofs, whever it originates.

this market-bounded model aligns directly with ethereum's long-term scaling roadmap. scaling the L1 without sacrificing decentralization requires a dramatic increase in zk proof generation. this creates a structural, non-speculative demand for a new commodity market: prover networks.

PROVE is the native currency for this market, creating a self-reinforcing flywheel:

as eth scales -> proof demand rises -> value accrues to PROVE -> which incentivizes a more robust and efficient prover network -> enabling larger computational workloads and further scaling.

ultimately, PROVE is positioned not as a bet on a single application, but as a stake in the foundational infrastructure for the entire onchain economy.

succinct is building the proving economy. PROVE is how you own it.

[crashout @0xCRASHOUT]

ethereum's endgame: scale the L1 to billions and L2s to hit trillions (of gas)

the tech barrier is no longer theoretical. SP1 Hypercube has shown real-time zkVMs are viable in production.

the EF plans to shift the L1 from block re-execution to ZK proof verification. this allows for:
- ~10,000 TPS on Ethereum
- increased gas block targets to gigagas (billions of gas/sec)
- increased blob throughput (DAS), lowering DA cost and letting rollups post more data per slot

this unlocks the full potential of L2s, targeting an aggregate L2 execution towards teragas (trillions of gas/sec) and ~10,000,000 TPS across the L2s.

the bottleneck shifts from validator hardware to DA bandwidth. we get maximum decentralization, security, and scability from ZK, leaving week-long challenge periods behind.

at teragas scale, even a miniscule price (e.g., ~$0.00002/proof) implies multi-billion $/yr gross L2 proving; a 10–15% take-rate sends hundreds of millions $/yr to the Succinct Prover Network (+ separate L1 proving revenue), shared with provers and PROVE stakers.

gprove.

[Succinct @SuccinctLabs]

looking sharp spreading the gospel of real-time proving at Frontiers @jtguibas

[Succinct @SuccinctLabs]

Succinct Mainnet Livestream https://t.co/LORI74TCRl

[Succinct @SuccinctLabs]

See you in 1 hour!

[Succinct Foundation @SuccinctFND]

The Succinct Prover Network is live on mainnet.

Today, we PROVE them wrong.

[Succinct @SuccinctLabs]

Tomorrow.

[Succinct @SuccinctLabs]

The moment you've been waiting for is finally here.

[The Kael | (❖,❖) @mrbuu25210652]

So, the time for another important stage is approaching. And the past two weeks have been truly crazy and vibrant. I got to participate in fun and exciting contests throughout this period and almost forgot I needed to prepare something for the big thing coming up. 🫡

Thank you, Master @_twolight , for collaborating with me on the idea for this this piece. This artwork celebrates the most passionate members of the @SuccantVN community, with a style that embodies Vietnamese culture, which we want to introduce to everyone. 🤝

A big thank you to the members who dedicated themselves fully to this recent journey. There are many more members, but I couldn’t include everyone in one piece of art, so I look forward to featuring you in future artworks. 🫶

The @SuccinctLabs summer is still here, and its real heat is about to spread more intensely than ever!! 🌞

@0xCRASHOUT
@adoins
@calicocat2025
@TeeHuc269
@0xhirono
@hooleesheetJack
@Liu_xoke
@linh_ricky
@Miley_nguyen2k
@ManekiNeko2686
@Hu14951738
#RoadToMainnet #Succinct
#SuccinctSummer

[Liuseco @Liu_xoke]

@0xSunbin @MCuongSuccinct @itzchi__ @advaith @Miley_nguyen2k @COOCOOWAVESS @Hu14951738 TYSM my friend

[Liuseco @Liu_xoke]

@Hedvigmolly @SuccinctLabs Nice choices broski

[Liuseco @Liu_xoke]

@Donald153Jack @SuccinctLabs Haha, your video so funny bro