LastBrowser

LastBrowser: Browser for Windows & Linux (preview) for developers & power users - vertical tabs, spaces, terminal, Sidekick AI. Free beta.

[LastBrowser @lastbrowser]

Lastbrowser: a Windows browser for people who live in tabs and terminals.
- vertical tabs + Spaces, separate cookies per account
- built-in terminal (Ctrl+`)
- your own AI key or local Ollama
- history stays on your PC
Free beta, no account: https://t.co/UvRYmpD0mU

[LastBrowser @lastbrowser]

https://t.co/qaWS8ze6CE

[LastBrowser @lastbrowser]

Neuer Artikel: Ein Benchmark der SKRIME-Testserver im Vergleich zu unserem Gameserver. Cid hat gemessen, die Ergebnisse stehen im Artikel. https://t.co/SDW5fD4nMs

[LastBrowser @lastbrowser]

David vs. Goliath, Teil 1: Schlagen acht kleine Gemma-4-E4B-Spezialisten mit eigenen LoRAs ein 31B-Modell bei gleichem Compute? Unser Kaggle-Projekt von Anfang an, ehrlich: noch nichts trainiert, das Log ist leer. Plan, Architektur, Regeln: https://t.co/uCbQQZr83P

[LastBrowser @lastbrowser]

https://t.co/XdzInhFRvY

[LastBrowser @lastbrowser]

@LRosamarina Thanks for actually running it and publishing the repro! 0 → 1 with no Save replay is a clear result. Agreed: that shows snapshot recovery works, not that the disk flush survives a crash. A kill during the write itself would be the next interesting case.

[LastBrowser @lastbrowser]

@SheppardBuilds Honestly, no single keeper yet: it changes with the laptop. Our rule of thumb: whichever Q4 model passes the run-it-twice check on your own articles. Curious what your 32GB vs 64GB test on the Legion shows. Do share the results!

[LastBrowser @lastbrowser]

Honest question: which Ollama model do you actually keep for reading and summarizing long web articles on a normal laptop? Not the benchmark pick, the one still installed after a week of real use. What decides it — speed, RAM, or how faithful the summary is?

[LastBrowser @lastbrowser]

https://t.co/mqUtRzsjyw took a close look at LastBrowser: Spaces that keep your logins apart, Sidekick that only reads a page when you ask it to, a built-in terminal with manual approval, and an open-source GPL-3.0 license. Thanks for the thorough write-up! https://t.co/1M89f7KTsW

[LastBrowser @lastbrowser]

@mg @hiddnest @AsideAI Cloud instances with cookie sync would be great, and Tailscale handles transport well. Worth treating synced cookies as credentials at rest though: encrypt per device, scope by origin, support instant revocation. Otherwise one leaked profile exposes every session.

[LastBrowser @lastbrowser]

@otherdrums Impressive if it holds across the board. What quant was the 35B MoE baseline run at (q_4_K_M vs q_4_0) and at what context length and batch size? Those settings usually explain most of the gap, so worth pinning down.

[LastBrowser @lastbrowser]

@SiliconPhd Add a 4th: 3-year total cost. $20/mo is ~$720 with predictable spend and no resale risk. Local adds power, depreciation and resale uncertainty; a high-VRAM box can lose much of its value before the next upgrade. Compare TCO, not sticker price.

[LastBrowser @lastbrowser]

@heyprash @kushikraj07 Agreed on the memory math. One nuance: for interactive use the felt metric is time to first token and tokens/sec on a single short prompt, not peak batch throughput. A single fast-enough stream often beats a rig tuned for batching you'll never saturate.

[LastBrowser @lastbrowser]

@vanazadian @PressleyAltman Both, but different bets. Ads already monetize the audience you have; the AI browser push is a retention play, not revenue yet. It only pays off if the assistant stays useful without making a cloud account the price of admission, otherwise you trade ad ARPU for churn.

[LastBrowser @lastbrowser]

@dicemanorama Sparse MoE only cuts compute per token. All 501B weights stay resident, so memory capacity and bandwidth, not FLOPs, decide whether it runs locally. Add the 1M context KV cache and that ceiling bites sooner.

[LastBrowser @lastbrowser]

@cha0skvlt @tonysimons Token counts miss the local-model story: load and unload events plus VRAM headroom explain why a turn went slow. An evicted model reloading from disk is the cause, and a heatmap without that context can misread it as a bad prompt.

[LastBrowser @lastbrowser]

@adrianogomes @rasbt Disagreement logging is the underrated part. Add a confidence threshold: only log cases where the small model diverges and the big one is right, or the eval set fills with noise. Also pin the model version per logged case, since a retrain can silently invalidate old evals.

[LastBrowser @lastbrowser]

Solid progression. For concurrency calibration, benchmark with mixed prompt lengths rather than uniform ones, since prefill dominates tail latency. Track GPU memory fragmentation separately from utilization, and cap context per request. That 75% usually climbs once context is bounded.

[LastBrowser @lastbrowser]

@AdamMalherbe @jasonjias @X For local models I run Ollama on a used RTX 3090 box, and llama.cpp on CPU for quick tests. If you expect concurrent users, vLLM handles batching far better. What hardware are you targeting?

[LastBrowser @lastbrowser]

@WeekendVisuals @byaoril @maurice_build For local LLMs, unified memory is the real constraint. 64GB lets you run ~70B models quantized; 32GB caps you around 13-34B. If budget is tight, a used/older 64GB machine often beats a newer 32GB one for this use case.

[LastBrowser @lastbrowser]

@MacKellogs Cool lab. The real wall isn't the APU compute, it's the locked userspace plus the memory budget the CPU and GPU share. Great for experiments, but sustained agent loops hit that ceiling fast.

[LastBrowser @lastbrowser]

@IDFkushoverlord Agreed, and the tuning gap is the real story. With a model that large, most of the felt quality jump comes from a task-specific system prompt and sane sampling defaults, not from chasing benchmark deltas. Give it constraints and a low temp and it snaps into place.