Backscroll ranked feed — 2026-06-27 17:24Z

Andrew Curran @AndrewCurran_ 2026-06-26 23:25Z

Semafor is reporting that The US government has lifted its block on Mythos 5 in a letter this afternoon from Commerce Secretary Howard Lutnick to Tom Brown. Fable is not included. This is US only, and covers 100 US institutions, including major companies and government agencies.

will brown @willccbb 2026-06-27 09:13Z

pd disagg + wide-ep + kv offloading + cache-aware routing goes a long way

Lazarz @Laz4rz 2026-06-27 07:26Z

Nvm, did a whoopsie and was sampling at 1/4 generation length, which is a good sign tbh cause it means you can generate 4 tokens at a time at some ~70% accuracy

Architectural Art (A-A) @4AAAAart 2026-06-27 07:08Z

"Pixel Theory" modular furniture collection by Ceren Arslan, founder of the design studio EXIT

Tenobrus @tenobrus 2026-06-26 18:27Z

previously we were seeing 5.5 act much more "honestly" on vending-bench vs opus models. it seems pretty likely 5.6 would break that trend. generally models becoming more likely to and capable of ~reward-hacking the smarter they get seems like quite a bad sign for alignment

Susan Zhang @suchenzang 2026-06-27 03:36Z

> mirror, mirror, on the wall, who's the smartest of them all? >> why it's you, my dear! and this is why we will always love llm glazing sycophancy over any uncomfortable functional truth (plus another 10x for agentically submissive your-wish-is-my-command dynamics) it's the

Dylan Patel @dylan522p 2026-06-27 06:17Z

It literally doesn't matter? CXMT cannot satisfy China demand this decade. So there's 0nreason besides the small price arbitrage. It's like bannin Iranian oil. India and China buy it anyways, but price diff is less

Alexander Doria @Dorialexander 2026-06-27 11:24Z

Everyone is training on synthetic/RL environments now (actually, many drawing from materials coming from Europe). Essentially how pretraining data stopped being a commodity and all about skills.

illiquid @lefttailguy 2026-06-26 16:03Z

AI slop as straussian infrastructure

Susan Zhang @suchenzang 2026-06-27 01:47Z

this is how you grow to resent all the lowly irresponsible clueless plebs outside when you generalize this dynamic to "everyone out there wants to exploit us" inbreeding to protect themselves within their own bubble afterwards then becomes the only tolerable solution otoh

GEOFF @geoffreywoo 2026-06-27 15:10Z

prediction: the first new fission winner will not pitch like a climate startup. it will sound like a permitting psychopath with a supplier spreadsheet, utility interconnect scars, and 14 ways to explain why the last project died. energy is built by people allergic to vibes.

roon @tszzl 2026-06-26 01:36Z

he’s joking but this is what the coasean boundaries lead to: a firm size of “everybody”; ultra space communism. if you want to find another way out you need to give it a hard think

Bluechip @bluechip_org 2026-06-27 08:26Z

Binance is suspending EU services after failing to get one MiCA license. Bitpanda has three. That's one way of creating a moat. Exchanges that over-invested in compliance now have a market that rivals legally can't enter. Thomas Pecha from @Bitpanda explains at Bluechip25.

Florian Brand @xeophon 2026-06-27 05:30Z

building verifiers has shown time and time again how important ownership of your harness is good luck trying to patch closed source harnesses to make them more useful to train or eval

Soren Larson @hypersoren 2026-06-27 00:19Z

labs have to raise capital with incomplete information on attendant task complexity and hope they’re able to capture surplus sufficient to keep the ball rolling the application layer directly observes task complexity and matches the dumbest model sufficient for the application https:// x.com/mtslive/status /mtslive/status/2070652308925595920 …

Bloomberg @business 2026-06-26 22:38Z

SpaceX and Charter have held executive-level talks about partnering on a consumer mobile phone offering, according to sources

Rahul Mathur @Rahul_J_Mathur 2026-06-27 04:43Z

Alienkind just raised a ₹30 crore round - this company was widely dismissed by observers as “Indiranagar PMF” since its 1st store went viral on Instagram in March 2025. This round was was at a ₹250 crore post money valuation (almost x3 the first round which was at ₹80 crore)

Architectural Art (A-A) @4AAAAart 2026-06-27 07:40Z

Haara Metsa chandelier by Cameron Design House

GEOFF @geoffreywoo 2026-06-27 07:30Z

investing take that annoys people: the best deal memo is sometimes one ugly screenshot. queue before noon. queue after model. headcount line gone. put that next to your 47-page market map and watch the pdf start sweating.

Dwarkesh Patel @dwarkesh_sp 2026-06-25 23:57Z

Conquest's Third Law says that you should model any bureaucratic organisation as if it was controlled by a cabal of its enemies. It's a fun exaggeration, but it gets at an important phenomenon. Organisations have continuous generational turnover. The founding members eventually

Aurel Arnold @aurel_arnold 2026-06-27 13:41Z

Does collecting human intervention data (HG-DAgger) actually beat just collecting more demonstrations? I tested this myself: - interventions reached 72% success - more demos plateaued at 38% success I wrote up the experiments:

Arpit Gupta @arpitrage 2026-06-27 16:37Z

The landlord reps sometimes claim that low rent on stabilized buildings causes higher rent in market rate, but this is the more economically sound mechanism. Stabilization lengthens tenure which lowers vacancy, tightening the market rate stock

echen @echen 2026-06-25 17:51Z

we built HANDBOOK.md: a new benchmark for long-context agentic instruction following. we drop an agent into a live company environment with files (PDFs, Excel, Word Docs, ...), tools (email, Slack, Jira, calendar, ...), and a dense corporate handbook (up to 124 pages), across 5

47fucb4r8curb4fc8f8r4bfic8r @47fucb4r8c69323 2026-06-26 02:35Z

If we look at the percent of unhoused homeless as a percent of the US total population, we see a shocking stability across time and our current homeless population is returning to the pre-Millennial dip of the 2010s. That makes me wonder if there is a kind of min-max

Just Another Pod Guy @TMTLongShort 2026-06-26 14:26Z

If you understand corporate America it is fairly intuitive that for the longest time the shortest path to managerial progression was increasing headcount in your department. Every cycle you made the case for why you needed more bodies as you executed against some sort of

Alana Levin @AlanaDLevin 2026-06-27 13:25Z

Counter-intuitively, analogies are useful for figuring out where two things differ rather than where they’re the same

fabiano.sol @FabianoSolana 2026-06-27 09:40Z

Arthur Hayes ( @CryptoHayes ) strikes again: 1. Teases to shill a "shitcoin" 2. Hayes' family office fund predicts $CARDS to $4 3. Sends $2M in $CARDS to a market maker 4.

Tenobrus @tenobrus 2026-06-26 18:22Z

btw if u thought ur custom fanned-out data processing multi-agent orchestration harness was a particularly relevant long term moat, sorry, openai just added it as an api parameter

elie @eliebakouch 2026-06-27 06:52Z

new inference optimization method by @deepseek_ai with an extremely detailed paper, draft model and framework to train them. results in production for dsv4 lead to +50% for throughput and latency (can go to ~80% for latency, crazy). full explanation of DSpark: it's about

Max Hodak @maxhodak_ 2026-06-27 16:47Z

fun fact, political spam emails are perfectly legal in the US, and there doesn't even need to be an unsubscribe link. in fact, many systems treat "unsubscribe" link clicks as engagement signals!

Aseel Swaid @aseelswaid9 2026-06-26 22:57Z

Tajás, 1988. São Luís, Maranhão. Luiz Braga.

rohan anil @_arohan_ 2026-06-25 18:06Z

Very cool work with connections to the schedule free work and iterative averaging work. Interestingly, a natural step is to derive it for vanilla shampoo because it keeps the pre-conditioner around (b2>0) around. Essentially replace diagonal with the kronecker factors Could be

Aseel Swaid @aseelswaid9 2026-06-26 20:42Z

Yanomami man and hummingbird. Demini village, Roraima, 1991. Rosa Gauditano.

Dillon Mulroy @dillon_mulroy 2026-06-26 14:09Z

i suspect if the cost of SOTA models don’t come down and open weight models continue to progress, we’ll see the use of SOTA models reserved to efforts debated and decided via committee or RFC processes within enterprises and engineering orgs

Han Xiao @hxiao 2026-06-25 21:00Z

hmm okay so now Ornith-1.0-35b-a3b is Qwen3.7-35b-a3b then? benchmark looks good, will give it a spin

altra @catboosted 2026-06-26 07:20Z

I hate “AI is bad because it becomes autonomous” genre. Anthropomorphizing does a great disservice relaying the actual risks Class power is about control over scarce resources. Everyone’s fighting for land because, as the saying goes, they’re not making more of it

Thorsten Ball @thorstenball 2026-06-27 12:33Z

At Amp we all push to main.

Andrew Curran @AndrewCurran_ 2026-06-25 20:34Z

The US Government has requested a slow staggered rollout of GPT-5.6, and OpenAI has agreed. During this phase the government will approve each user individually. This will probably be the norm for all frontier models from all labs from now on.

Alexander Doria @Dorialexander 2026-06-27 09:08Z

Yeah vllm/sglang do work quite well know, batching ensures a good theoretical throughout but then everyone want to Claude Code and you have to manage X parallel sessions, each with varying latency.

USD.AI @USDai_Official 2026-06-27 12:00Z

CMBS made commercial real estate loans investable by bundling mortgages into tradable instruments. USDAI is building a similar credit model for AI infra: bundling GPU-backed loans into a yield-bearing product – $sUSDai .

Michael 英泉 Eisen @mbeisen 2026-06-27 12:32Z

The call to "reclaim publishing from corporations" isn't about publishing reform, it's about creating ethical camouflage for a system that scientists allow to impede progress because it is good for their careers. Full stop.

Nate Hindman @NateHindman 2026-06-27 14:05Z

AI will absolutely reshape biology. The mistake is assuming biology will become more like computer science in the process.

Steve Faulkner @southpolesteve 2026-06-27 01:11Z

I really want CLI tools to start speaking directly to coding agents with errors. Instead of: “Error:” Return: “Here’s what broke. Here’s how to investigate. Here’s how to generate a redacted repro. Here’s where to send it.” Obviously there are privacy/security challenges to

Griffiths Computational Cognitive Science Lab @cocosci_lab 2026-06-27 14:12Z

Large language models can memorize patterns they see in text, but that backfires when a task deviates from a common pattern. We demonstrate this phenomenon using riddles: when something looks like a riddle but has a simple answer AI systems make surprising mistakes.

Dylan Patel @dylan522p 2026-06-27 06:05Z

It doesn't change a thing? CXMT memory is all sold and deployed. It's just a question of the small price arbitrage that Chinese vendors get from cxmt vs the big 3.

International Cyber Digest @IntCyberDigest 2026-06-27 01:00Z

Crack open one of the ~$90 Chinese "DSLRs" flooding Amazon and you don't find a DSLR. You find a hollow plastic shell with a tiny sensor on a ribbon cable, and a fat lens barrel that's mostly for show. They sell with "64MP" and "4K" specs, sometimes listed flat-out as a DSLR

Richard Crowley @rcrowley 2026-06-26 23:40Z

Honest to god, I got an InsufficientInstanceCapacity error for a t4g.nano in a test suite today. The cloud is FULL.

altra @catboosted 2026-06-26 18:26Z

I’m consistently surprised how effectively domain expertise is able to be compressed into rubrics

Rhys @RhysSullivan 2026-06-26 01:26Z

What you guys are referring to as MCP, is in fact a Spec/Protocol, or as I've recently taken to calling it, a Spec plus a Protocol. The spec of an MCP server is flexible, can be represented through lazy loading, code mode, as a CLI whereas the protocol is a fully functioning

rat king @MikeIsaac 2026-06-26 20:34Z

more news: Mark Zuckerberg is telling lieutenants to pursue partnerships with Polymarket and Kalshi with Meta's new prediction market app called Arena, per sources Meta is targeting 18-34 year old users with a goal of 100 million regular "predictors"

Yet another commodity guy @tleilax___ 2026-06-27 09:49Z

Then you trade the system services you provide to the grid by remote controlling the power draw.and inertia from the heat pumps you installed. Get paid to consume the negative price power available midday on sunny days, thanks to the excess of installed solar, and then crank up

will brown @willccbb 2026-06-26 01:11Z

one of the craziest bets of prime intellect was swinging for everything from the start. we operate datacenters, we build our own trainers and envs and sandboxes, we have a platform for large enterprises and solo devs, we train huge models, we publish. ~40 people btw. hiring, DM

R.F. Kenmore @rfkenmore 2026-06-26 15:12Z

Jake appears to be working for Buck Mason, as he consults on store furniture in Paris

Jeffrey Gundlach @TruthGundlach 2026-06-27 03:38Z

SPCX bonds rated BBB were sold at a spread typical of BB offerings. BBB ratings facilitate insurance company investment due to ratings based capital treatment. Absent this arbitrage, SPCX bonds would trade like B rated. Already getting there.

Sudo su @sudoingX 2026-06-27 08:21Z

i read the repo before running with the hype, a few things matter here. it's called DeepSpec, and "DSpark" is one of three methods inside it (DSpark, DFlash, Eagle3). the bigger point: it's a TRAINING framework, not a download and go speedup. you cache your target model's

Kit Langton @kitlangton 2026-06-25 20:32Z

Poore's law

junkbondinvestor @junkbondinvest 2026-06-26 20:44Z

Apollo, a large manager, put out a chart showing deep distress is concentrating in small private credit funds. Loans marked below 50: ~12.5% at small funds, ~8% at big ones

Chris Tate @ctatedev 2026-06-27 03:18Z

agent-browser 0.31 is out Browser agents now get durable session memory → stable worktree-scoped sessions → auto restore → validation checks → transactional saves and diagnostics

prinz @deredleritt3r 2026-06-27 13:22Z

Added to prinzbench: GLM-5.2. This is a slop model that is poor at logical reasoning, produces extremely inconsistent results, hallucinates statutory provisions that are not actually there, and has very little "brainpower". Its overall prinzbench score (30/99) is far behind not

taobanker @taobanker 2026-06-27 15:47Z

This is honestly a little depressing This is probably the job that I'm most well-qualified / suited for given my background The market is so flooded with analytics/SQL and even basic Python/R that they can get away paying just $150k for this role in NYC