Backscroll ranked feed — 2026-07-03 20:55Z

Susan Zhang @suchenzang 2026-07-03 13:52Z

books3 was absolutely necessary for llama to happen back in the day. you can even find my pathetic cleanup code for it here: https:// github.com/facebookresear ch/metaseq/tree/main/preprocessing/books3 … it was also noted to be a critical datasource for mmlu in the llama1 report, though by now we know mmlu is useless and high

Andrew Curran @AndrewCurran_ 2026-07-03 15:28Z

Alibaba has banned employees from using Claude Code, they are being told to use the company's own coding platform Qoder.

Tenobrus @tenobrus 2026-07-03 19:49Z

personally i take accounts who post under their name and face much less seriously until proven otherwise. it's a p strong single of "linkedIn-intent" and i don't really give a shit about personal capital building human-slop

47fucb4r8curb4fc8f8r4bfic8r @47fucb4r8c69323 2026-07-03 19:36Z

Poetry is manipulating natural language to produce an experience. Coding is manipulating natural language to produce an experience. Obviously the skills and temperament to be good in one are different than those required for the other. But understood as ways people use

Aseel Swaid @aseelswaid9 2026-07-03 19:34Z

Andrew Curran @AndrewCurran_ 2026-07-03 18:37Z

Uno Reverse! From the Midjourney motion: 'If Plaintiffs are developing image-generating AI models — trained on unlicensed, third-party copyrighted data — for internal use in storyboarding or ideating content for film or TV, that evidence would equally demonstrate that it is an

Nathan Flurry @NathanFlurry 2026-07-03 19:57Z

tldr why i'm excited for executor/integrations: → everyone need a "company brain" → everyone thinks memory is the issue → but nobody gets past the integrations phase → tackle integrations first, everything else comes later and most importantly oss + self-hostable

Jakub Krehel @jakubkrehel 2026-07-03 14:50Z

Some of my favorite parts of the new Interfere product pages. Every asset is fully built in code. The animations are built with Motion, except for the dot matrix, which is uses a canvas. Had a ton of fun working on these with @paulfaivret .

Applied Compute @appliedcompute 2026-07-03 18:59Z

Our formula takes in six inputs: train throughput, sampling throughput, training batch size, queue capacity, rollout concurrency, and response-length tailness. We validated our discrete-event simulations against real Qwen3-8B GRPO training runs, resulting in predictions that

Zuhayeer Musa @zuhayeer 2026-07-03 18:59Z

We pulled data to compare software engineer pay and hardware engineer pay at the chip companies. Median hardware pay beats median software pay at 7 out of the 10. Nvidia is the only meaningful example where software still beats out hardware.

∿ @somewheresy 2026-07-03 18:54Z

The datacenter backlash is best read as a land-use and fiscal-allocation conflict: affluent suburban and exurban homeowners (historically the net beneficiaries of a structural subsidy flowing from denser, more tax-productive land) are reacting to public investment (utility capex,

Nikunj Kothari @nikunj 2026-07-03 14:28Z

I tell every founder to have a 3 things you need to believe in to invest up top and force the investor to commit in a binary way. Usually saves them a lot of time.

Adam Karvonen @a_karvonen 2026-07-03 14:38Z

Also, empirically today's open models dominate the cheaper side of the Pareto frontier. Qwen-3.6-35B-A3B is smarter than haiku 4.5 and 10x cheaper for example.

wavefnx @wavefnx 2026-07-03 19:51Z

Apollonian, but Hermetic, but Apollonian Strawberry Full Moon (June's Moon) behind the temple of Poseidon

Susan Zhang @suchenzang 2026-07-03 14:55Z

leaky paris-london office, apparently or at least whatever illusion of leaks there were to give the llama1 authors the audacity to write that

Alexander Doria @Dorialexander 2026-07-03 09:45Z

>indefinite compute >likely largest model ever >good arch research (and copying deepseek is free) >come on, do your agent thing >… what did you try to model exactly, anon?

Jake Castillo @jakecastilloooo 2026-07-03 18:44Z

This is technically correct, but it’s not the whole picture. Because if you run this test 100x with ads, meta gives you roughly the same result 100x. But with UGC, 30% of the time you’ll get those crazy asymmetrical upside results, and 70% you’ll get low UGC views/results.

Junxian He @junxian_he 2026-07-03 16:54Z

We appreciate many model developers to adopt Toolathlon to evaluate models' personal intelligence on diverse tool calling in complex workflows. Since Toolathlon's first release, some live tasks are drifting from the original groundtruth, some tasks are found to have flaws on the

47fucb4r8curb4fc8f8r4bfic8r @47fucb4r8c69323 2026-07-03 12:00Z

One of the hard things to understand about Wittgenstein is that his view of logic deepened his faith in the metaphysical. By showing the boundaries of logic, he showed the necessary existence of that which is beyond the boundary of logic; and so, arguing any metaphysical point

Andrew Curran @AndrewCurran_ 2026-07-02 20:54Z

On the heels of reports that META is exploring a move into compute-as-a-service like xAI, Mark Zuckerberg told an internal town hall that AI agent development over the last four months hasn’t accelerated 'in the way we expected'. The race continues to narrow.

Lazarz @Laz4rz 2026-07-03 13:49Z

ok @OpenAI what did you do to 5.5? this isnt normal

Erik Bernhardsson @bernhardsson 2026-07-02 23:11Z

Area around Yakutsk in Siberia

Architectural Art (A-A) @4AAAAart 2026-07-03 06:33Z

Milanese apartment renovated by Atelierzero Milan, Italy

Minh Nhat Nguyen @menhguin 2026-07-03 18:16Z

nobody: the claude computer use model retrying a dead archived link necessary to complete the task for 600 turns before reward hacking the grader:

Deirdre Bosa @dee_bosa 2026-07-02 17:27Z

Alternative take: the US taking 5% of OAI makes sense. Its a dividend on the public conditions that made the company possible. And China didn’t dominate EVs, batteries, solar, drones, telecom by pretending strategic tech is just software Industrial strategy with upside

Guillermo Rauch @rauchg 2026-07-03 19:50Z

Agentic self-improvement. Give your agent the ability to introspect its past runs, spot inefficiencies, errors, redundant tool calls, and produce new prompts and skills. That’s why agent observability is built-in when you deploy http:// eve.dev to Vercel.

nishant modi @nishantmodi3105 2026-07-03 18:19Z

if ur loops retry on the same queries, check ur memory layer before touching ur agents. i swapped ChromaDB for Graphiti + Neo4j (temporal knowledge graph): results: staleness error: 87% → 20% avg latency: 11,937ms → 9,243ms full writeup: https:// nishantmodi.me/posts/temporal -memory-architecture.html …

alpertunga.eth @alpertungaakkus 2026-07-03 17:33Z

#SSAAT faize atmak için halktan para topluyor... İzahnameyi okuyunca tablo ilginç: • Halka arz gelirinin %90'ı işletme sermayesine ayrılıyor. • Kullanılana kadar para vadeli mevduatta nemalandırılacak. • Yatırıma ayrılan pay ise sadece %10. Bir de üzerine değerlemeye

1776 Live @250YearsAgoLive 2026-07-03 15:48Z

Independence may be on hold, despite yesterday’s decision. Continental Congress has been meeting to further consider the Declaration of Independence, which they have hastily revised to remove author Thomas Jefferson’s denouncement of the transatlantic slave trade. More changes

forward deployed ccp gf @FangYi11101 2026-07-03 14:06Z

Netflix is also doing smaller scale team-based “layoffs” as of late btw

Aseel Swaid @aseelswaid9 2026-07-03 19:20Z

Michael @endpointarena 2026-07-02 18:59Z

Endpoint Index separates clinical trial risk by "sleeve", so you can track where sentiment is rising or falling across immunology, oncology, etc

Yohei Nishitsuji @YoheiNishitsuji 2026-07-03 13:30Z

#つぶやきGLSL float i,e,R,s;vec3 q,p,d=vec3(FC.xy/r*.8+vec2(-.4,.6),1);for(q.zy--;i++<99.;){o.rgb+=.02-hsv(s*s+R*4.,.4,min(e,R)/6.);s=12.,p=q+=d*e*R*.1;p=vec3(log2(R=length(p))-t*.4,exp(R-p.z*.5),atan(p.y*.6,p.x)+t*.2);for(e=--p.y;s<3e2;s+=s)e+=dot(sin(p.yzy*s),sin(p.xxz*s))/s;}

David K @DavidKPiano 2026-07-03 14:57Z

The more complex the agents I build, the more state machines (statecharts) feel like the right artifact. Run the same inputs through different machines & compare. Keep the winner, or eval at transition level & combine the good parts. Feels like building agents w/LEGO.

sophia @cis_female 2026-07-03 15:16Z

similarly to what @gabo_de_cuba says apparently these were both just standard rates for "intermediate countries" under the "OECD consensus" categories. Spanish interbank rates were 15-18% in that timeframe. 1980 order: https:// boe.es/buscar/doc.php ?id=BOE-A-1980-12595 … 1982 (17681) https:// boe.es/boe/dias/1982/ 07/16/pdfs/A19298-19299.pdf …

Maxim Enis @maxim_enis 2026-07-03 19:10Z

you should probably be recording your computer screen at all times so that future LLMs have all the context needed to automate your work

deep Manifold @BetaTomorrow 2026-07-03 12:30Z

#DeepManifoldInterpretation Title: EdgeBench: Unveiling Scaling Laws of Learning from Real-World Environments Author: ByteDance Seed 1-linear: pay attention to the fixed point field. The fixed-point field exists as the broader space of possible fixed-point classes, basins, and

Gizem Akdag @gizakdag 2026-07-03 19:32Z

Postcards from tennis courts

InspirationGx @inspiration_gx 2026-07-03 17:05Z

I built time-shredr for my @solanaturbine Capstone an ebpf program that hooks up to uprobe events inside the PoH component of solana Validators. see ticks, record, record_batches, hash grinding etc as it happens inside the validator. for this demo tested my hookpoints on the

GEOFF @geoffreywoo 2026-07-03 09:50Z

deep tech founders lose me when every risk is “execution.” execution is not a risk category. it is hiding the risk. name the demon: thermal drift oxide contamination vacuum seal failure packaging yield one supplier in ohio with a 38 week lead time

James Cowling @jamesacowling 2026-07-03 17:56Z

A common problem on large projects is solving problems on the wrong layer of the stack because you’re afraid to change the layer below. e.g., adding checks and retries because you don’t know why the server code isn’t transactional. This is how simplicity dies and LLMs will very

Dan Gray @credistick 2026-07-03 14:06Z

What matters is the multiple from current round to next round, and current price isn't really useful information to understand that. Ultimately, people just want a market signal to assure them that they have made a good choice. Price is a dumb signal, but it's an obvious one.

Kat ⊷ the Poet Engineer @poetengineer__ 2026-07-03 17:59Z

sizing each fork's possibility-space volume: i embed the options, and the cube = how far apart they spread in meaning-space. four near-identical tweaks = tiny cube, genuinely different directions = bigger. the path is the project over ~2 months, spiraling bottom to top in order.

bruno @tvnxty 2026-07-03 15:38Z

I remade sound design for @tembo from scratch! Sound on! Built from the ground up, more energy, more cohesion and audio that actually flows with the visuals, not as a secondary, but as an equal. Visual overhauls should be exciting, and what better way to make them so, than

Applied Compute @appliedcompute 2026-07-03 18:59Z

Controlling staleness in an async RL stack has not been well understood, so we derived a closed-form formula that predicts staleness in advance. Our predictions match measured staleness from production RL runs within a fraction of a step. Building these simulations has led to key

Lucas Beyer (bl16) @giffmana 2026-07-02 19:07Z

Of course, love me some confirmation bias! DCLM was for text, DCVLM is the same for vision: analyzing VLM data mix across scales. Filtering is no bueno, but the "type" mixing matters, with unfortunately (but imo expectedly) the best small-scale mix != the best mid-scale mix.

Chris Tate @ctatedev 2026-07-03 16:55Z

Nuxt adapter has landed in emulate Stateful emulators for Vercel, GitHub, Google, Slack, Apple, Microsoft, Okta, AWS, Resend, Stripe, MongoDB Atlas, Clerk, Linear, Twilio Now with Nuxt support for same-origin

Eric Zhang @ekzhang1 2026-07-03 05:55Z

Something I'm pondering: Prefill and decode seem to have quite different dynamics in a disaggregated setup… on unexpected traffic spike, decode slows down a bit, but prefill just dies + drops requests it also seems harder to measure if prefill is close to gpu saturation

Thomas Schulz @thomasschulzz 2026-07-03 18:58Z

Struggling w influencers/UGC rn myself. Last few videos all have cleared hundreds of thousands/millions of views and literally getting sub 200 downloads a video

maja @majamediaco 2026-07-03 18:08Z

many companies could benefit from hiring a chief lore officer

Jongwon Park @ ICML @JongwonPar9958 2026-07-03 14:23Z

MacDiarmid et al showed reward hacking in narrow domain can generalize into alignment faking and sabotage. The fix they've deployed is inoculation prompting (2511.18397). But a newer paper (2604.25891) finds inoculation may not remove the misalignment so much as hide it. It

Byrne Hobart @ByrneHobart 2026-07-02 23:47Z

The doomers live a few model releases ahead of us, the it's-a-bubble crowd is a few models behind, only moderate AI bulls are truly living in the moment.

Design Reviewed | Graphic Design History @DesignReviewed 2026-07-03 15:32Z

Design Reviewed was created to give designers, researchers, and creatives affordable access to design history. The archive is free to explore online; all designer profiles, scanned artefacts, country and movement pages are publicly accessible. Membership from £2.49/month unlocks

John Arnold @johnarnold 2026-07-03 13:56Z

For nearly 50 years the standard of care for joint sprains has included rest, ice, and ibuprofen not bc rigorous evidence showed that worked, but bc that made intuitive sense and seemed to help. Now there's good evidence those actually impede recovery. Add it to a long list of

Andrew T. Walker @AndrewTWalker 2026-07-03 18:19Z

The rule of American Order is backed by the rule of American Force.

Jaytel @Jaytel 2026-07-03 13:09Z

i'm in the midst of purchasing a vitsoe unit. but hard to visualize from just the drawings they share fable 5: - researched + measured every 606 component then rebuilt the whole system as parametric USDZ - vision QA'd its renders in a loop until every component matched 1:1 -

Matthew Cassinelli @mattcassinelli 2026-07-03 13:52Z

If you're having performance issues with Siri AI, don't let App Intents be the bottleneck. Developers can use EntityCollection for large datasets – here's an example where tagging goes from 1,199ms to 4ms:

tomie @tomieinlove 2026-07-02 18:58Z

The future will hold such a superabundance of machine intelligence that we’ll use quadrillion-parameter models just to stream out the tokens for digital rain effects on our desktop wallpapers.

Silicon Jungle @JungleSilicon 2026-07-03 12:56Z

adding ai agents to slack makes it easy to work in parallel in separate threads. adding annotations to a browser harness brings the agents closer to the *work* itself.

Tzu-Heng (Brian) Huang ICML'26 @zihengh1 2026-07-03 12:25Z

Exactly! While individual programs can be limited by their expressivity, aggregating multiple of them can produce a better evaluator. Moreover, even if we route back to an LLM evaluator at the end, running programs as checkers first doesn't take much longer and gives us instant