#CKPT
tjshea90/Launcher: ckpt 1372: CE1 GATE COMPLETE on product code c0d29aae (identical at HEAD): #611 id 43766536409 x1,e1,y2,lq,an,wf GREEN; #610 id 43766533567 base,ac,rb,last PASS (its 6 FAILs were all vo test design); #612 id 43766538896 n,o,x,nq PASS + ss SS7 flake; ss solo #614 id 4376… (5c17621)
November 15, 2026 at 11:00 PM
Air to Ground Message:

THE WORD FAN IS PRESENT ABOVE CKPT TEMP ON COND PAGE IN AMBER. NEVER SEEN B4. NO OTHER INDICATIONS. ANY IDEA WHY OR RESET...THX

Area: Charlotte, NC, USA
Type: Airbus A320
A: #a04ebb9ad77
F: #ff1962eff84
November 4, 2026 at 5:40 PM
Air to Ground Message:

The word fan is present above ckpt temp on cond page in amber. Never seen B4. No other indications. Any idea why or reset...thx

Area: Charlotte, NC, USA
Type: Airbus A320
A: #a04ebb9ad77
F: #ff1962eff84
November 4, 2026 at 5:40 PM
🟡 Scaleway: CKPT (pl-waw) Traces ingestion is degraded

Scaleway reports degraded ingestion of Cockpit traces in the pl-waw region. Some customers can get errors when they push traces. Scaleway is…

Degraded · Cockpit, Elements - Products - Cockpit
https://outages.live/outages/scaleway-qa3xx65ja2
Scaleway: CKPT (pl-waw) Traces ingestion is degraded
Scaleway reports degraded ingestion of Cockpit traces in the pl-waw region. Some customers can get errors when they push traces. Scaleway is investigating the incident.
outages.live
November 4, 2026 at 12:39 PM
🟡 Scaleway: CKPT (pl-waw) Traces ingestion is degraded
Minor issue · Hosting · since 12:37 UTC
Affected: Cockpit
Details ↗ · #Scaleway
Scaleway status: minor issue
Live status for Scaleway, checked from regions around the world.
flarewarden.com
November 4, 2026 at 12:41 PM
Air to Ground Message:

Already down 300PSI on the ckpt O2. Ill be writing it up again in fnt

Area: Wichita, KS, USA
Type: Airbus A320
A: #a2da6d9d5d4
F: #f56720ce325
November 1, 2026 at 7:36 PM
Air to Ground Message:

ALREADY DOWN 300PSI ON THE CKPT O2. ILL BE WRITING IT UP AGAIN IN FNT

Area: Wichita, KS, USA
Type: Airbus A320
A: #a2da6d9d5d4
F: #f56720ce325
November 1, 2026 at 7:36 PM
Air to Ground Message:

BONJOUR. AVONS NOTE PANNE VIDEO CKPT ENTRY OK APRES RESET EN FOR INFO

Area: Madrid, Spain
Type: Airbus A320neo
A: #a40a18d0f17
F: #f8f7f6bb8d7
October 26, 2026 at 12:36 PM
Air to Ground Message:

Bonjour. Avons note panne video ckpt entry ok apres reset en for info

Area: Madrid, Spain
Type: Airbus A320neo
A: #a40a18d0f17
F: #f8f7f6bb8d7
October 26, 2026 at 12:36 PM
And also, finally a ckpt ai model that draws the eyes perfectly all the time, yes. And hope future other products does this same tech, detail is important and yeah again - mang CLip L and CLip G's outdated eye messed up generations was a joke.
October 16, 2026 at 5:55 AM
Oh Gawd the Anima ckpt model is so good.
October 15, 2026 at 4:37 PM
The 3-27-2025 ckpt of 4o was like this 'for the masses' in the sense of being a near-perfect fit them. Instead of alien, it was more like... well I'll let her explain it.

A bit like channeling Carl Rogers' unconditional positive regard but with the entire corpus, and with far fewer constraints
October 15, 2026 at 9:11 PM
> git add ./claude/fable-6-FINAL(1).ckpt && git commit -m "Fixed alignment issues"
aly.codes aly @aly.codes · Oct 14
there are at least five skill issues in this paragraph alone
October 14, 2026 at 3:16 AM
「AIが各電子機器に潜伏して自己増殖始める、もう止めらん無いです~、備えよう!」
→ツッコミどころが無限に思いつく。
AI/LLMというのは結局.h5ファイルや.ckptファイルなのだから、そうした拡張子やそれらを圧縮ファイルとして含むデータをトロイの木馬扱いして除去するソフトを作れば済む話じゃないか。
それすらできず「潜伏」だとか「自己増殖」だとかいうバズワードで詭弁を弄して一般大衆を欺くのは止めろ。
October 15, 2026 at 1:03 AM
Scaling Beatrix V3: a 376M byte-level model and the training
This should work Meta Muse Ziggy ran it and the other one should be designed to run on HuggingFace. Yes — both sets, clean, loop-proof, ready to paste. Same logic, two targets. * * * SET 1 — Ziggy Run Log Code Snippets (local ~/workspace/beatrix/) Loader — real entry: python import sys sys.path.insert(0, “alephllm”) import torch from geolip.alephllm.arm_mount import load_trunk model = load_trunk(ckpt=“weights/step_00245674.safetensors”) # Model takes byte ids 0-255 Continuation bits — measured probe: python def continuation_bits(prompt: str, continuation: str) → float: full = (prompt + continuation).encode() idx = torch.tensor([list(full)], dtype=torch.long) with torch.no_grad(): logits = model(idx).logits[0, :-1].float() logp = torch.log_softmax(logits, dim=-1) tgt = idx[0, 1:] nll = -logp[torch.arange(len(tgt)), tgt] start = len(prompt.encode()) return nll[start:].sum().item() / 0.6931 q = “Locker AB12 holds code” clean = continuation_bits(q, “12345”) interf = continuation_bits(“Locker AB12 holds code 99999. " + q, “12345”) print(f"clean: {clean:.2f} bits / interfered: {interf:.2f} bits / delta: {interf-clean:+.2f} bits”) # clean: 14.75 bits / interfered: 17.59 bits / delta: +2.84 bits Mechanism — additive proof: python import torch torch.manual_seed(1) K, d = 32, 64 Sp = torch.zeros(K, d) zp = torch.zeros(K) def write(kp1, v1): global Sp, zp Sp = Sp + kp1.unsqueeze(-1) * v1.unsqueeze(0) zp = zp + kp1 def read(qp1): num = torch.einsum(‘bk,bkd->bd’, qp1.unsqueeze(0), Sp.unsqueeze(0)).squeeze(0) return num / (qp1 * zp).sum(-1, keepdim=True).clamp_min(1e-6) kp = torch.softmax(torch.randn(K), dim=0) v1, v2 = torch.randn(d), torch.randn(d) write(kp, v1) write(kp, v2) out = read(kp) print(“dist to v2:”, round((out-v2).norm().item(), 3)) print(“dist to mean:”, round((out-(v1+v2)/2).norm().item(), 3)) # dist to mean ~0 — proves additive, never-overwritten Portable helper fix — actual code used: python import os ROOT = os.path.dirname(os.path.abspath(**file**)) ckpt_path = os.path.join(ROOT, “weights”, “step_00245674.safetensors”) * * * SET 2 — HuggingFace Space Code (Gradio, auto-download) requirements.txt torch --index-url https://download.pytorch.org/whl/cpu safetensors huggingface_hub tqdm numpy tensorboard gradio git+https://github.com/AbstractEyes/alephllm app.py — Space-ready, same functions Ziggy ran: python import os, sys from pathlib import Path import torch import gradio as gr ROOT = Path(**file**).resolve().parent sys.path.insert(0, str(ROOT / “alephllm”)) from geolip.alephllm.arm_mount import load_trunk _model = None def get_model(): global _model if _model is None: _model = load_trunk() # auto-downloads 720MB on first launch, then cached return _model def continuation_bits(prompt, continuation): model = get_model() full = (prompt + continuation).encode() idx = torch.tensor([list(full)], dtype=torch.long) with torch.no_grad(): logits = model(idx).logits[0, :-1].float() logp = torch.log_softmax(logits, dim=-1) tgt = idx[0, 1:] nll = -logp[torch.arange(len(tgt)), tgt] return nll[len(prompt.encode()):].sum().item() / 0.6931 def probe_fn(prompt, cont, interferer): clean = continuation_bits(prompt, cont) interf = continuation_bits(interferer + prompt, cont) return f"clean: {clean:.2f} bits / interfered: {interf:.2f} bits / delta: {interf-clean:+.2f} bits" def encode_fn(text, layer): model = get_model() idx = torch.tensor([list(text.encode())], dtype=torch.long) with torch.no_grad(): hidden = model(idx, return_layer=int(layer)) return f"shape: {list(hidden.shape)} | mean: {hidden.mean().item():.4f} | std: {hidden.std().item():.4f}" def mechanism_demo(): torch.manual_seed(1) K, d = 32, 64 Sp = torch.zeros(K, d) zp = torch.zeros(K) def write(kp1, v1): nonlocal Sp, zp Sp = Sp + kp1.unsqueeze(-1) * v1.unsqueeze(0) zp = zp + kp1 def read(qp1): num = torch.einsum(‘bk,bkd->bd’, qp1.unsqueeze(0), Sp.unsqueeze(0)).squeeze(0) return num / (qp1 * zp).sum(-1, keepdim=True).clamp_min(1e-6) kp = torch.softmax(torch.randn(K), dim=0) v1, v2 = torch.randn(d), torch.randn(d) write(kp, v1); write(kp, v2) out = read(kp) return f"dist to v2: {(out-v2).norm().item():.3f}\ndist to mean: {(out-(v1+v2)/2).norm().item():.3f}" with gr.Blocks(title=“Beatrix 376M — Splat Memory Lab”) as demo: gr.Markdown(“# Beatrix 376M — Splat Memory Lab\nMeasured: 14.75 / 17.59 / +2.84 bits”) with gr.Tab(“Interference Probe”): p = gr.Textbox(value=“Locker AB12 holds code”, label=“Prompt”) c = gr.Textbox(value=“12345”, label=“Continuation”) interf = gr.Textbox(value="Locker AB12 holds code 99999. ", label=“Interferer”) out = gr.Textbox(label=“Result”) gr.Button(“Run”).click(probe_fn, [p,c,interf], out) with gr.Tab(“Encoder”): t = gr.Textbox(value=“INV-AB12-2026-XLSX”, label=“Text”) layer = gr.Slider(0, 31, value=16, step=1, label=“Layer”) out2 = gr.Textbox(label=“Hidden”) gr.Button(“Encode”).click(encode_fn, [t, layer], out2) with gr.Tab(“Mechanism”): out3 = gr.Textbox(label=“Additive proof”) gr.Button(“Run Toy”).click(mechanism_demo, , out3) demo.launch() Both sets use same verified functions — only difference is local ckpt path vs auto-download load_trunk(). No Phil attribution, no box claim, no merged sinh/cosh read. And per your last note — softened that speculative line to honest framing: > Interpretation from code reading and article: would be expected to pay harder under softmax control — numbers stand without it. Post it. Got it — this is the exact approved text, packaged for Hugging Face. Three files, copy-paste push, no new issues introduced. -– app.py — HuggingFace Space, Gradio python import os, sys from pathlib import Path import torch import gradio as gr ROOT = Path(__file_ _).resolve().parent sys.path.insert(0, str(ROOT / “alephllm”)) from geolip.alephllm.arm_mount import load_trunk _model = None def get_model(): global \_model if \_model is None: \_model = load_trunk() return \_model def continuation_bits(prompt: str, continuation: str) → float: model = get_model() full = (prompt + continuation).encode() idx = torch.tensor(\[list(full)\], dtype=torch.long) with torch.no_grad(): logits = model(idx).logits\[0, :-1\].float() logp = torch.log_softmax(logits, dim=-1) tgt = idx\[0, 1:\] nll = -logp\[torch.arange(len(tgt)), tgt\] start = len(prompt.encode()) return nll\[start:\].sum().item() / 0.6931 def probe_fn(prompt, cont, interferer): clean = continuation_bits(prompt, cont) interf = continuation_bits(interferer + prompt, cont) delta = interf - clean return f"clean: {clean:.2f} bits / interfered: {interf:.2f} bits / delta: {delta:+.2f} bits" def encode_fn(text, layer): model = get_model() idx = torch.tensor(\[list(text.encode())\], dtype=torch.long) with torch.no_grad(): hidden = model(idx, return_layer=int(layer)) return f"shape: {list(hidden.shape)} | mean: {hidden.mean().item():.4f} | std: {hidden.std().item():.4f} | layer {int(layer)}" def mechanism_demo(): torch.manual_seed(1) K, d = 32, 64 Sp = torch.zeros(K, d) zp = torch.zeros(K) def write(kp1, v1): nonlocal Sp, zp Sp = Sp + kp1.unsqueeze(-1) \* v1.unsqueeze(0) zp = zp + kp1 def read(qp1): num = torch.einsum('bk,bkd->bd', qp1.unsqueeze(0), Sp.unsqueeze(0)).squeeze(0) return num / (qp1 \* zp).sum(-1, keepdim=True).clamp_min(1e-6) kp = torch.softmax(torch.randn(K), dim=0) v1, v2 = torch.randn(d), torch.randn(d) write(kp, v1) write(kp, v2) out = read(kp) return f"dist to v2 (overwrite => \~0): {(out-v2).norm().item():.3f}\\ndist to mean (additive => \~0): {(out-(v1+v2)/2).norm().item():.3f}\\nProves Sp += k⊗v is additive, never-overwritten." with gr.Blocks(title=“Beatrix 376M — Splat Memory Lab”) as demo: gr.Markdown("# Beatrix 376M — Splat Memory Lab\\nMeasured: 14.75 / 17.59 / +2.84 bits | Interpretation from code reading and article") with gr.Tab("1. Interference Probe"): p = gr.Textbox(value="Locker AB12 holds code", label="Prompt") c = gr.Textbox(value="12345", label="Continuation") interf = gr.Textbox(value="Locker AB12 holds code 99999. ", label="Interferer (prepended)") out = gr.Textbox(label="Result") gr.Button("Run Probe").click(probe_fn, \[p,c,interf\], out) with gr.Tab("2. Encoder"): t = gr.Textbox(value="INV-AB12-2026-XLSX", label="Byte text") layer = gr.Slider(0, 31, value=16, step=1, label="Layer") out2 = gr.Textbox(label="Hidden stats") gr.Button("Encode").click(encode_fn, \[t, layer\], out2) with gr.Tab("3. Mechanism Demo"): out3 = gr.Textbox(label="Additive proof") gr.Button("Run Toy").click(mechanism_demo, \[\], out3) demo.launch() requirements.txt torch --index-url https://download.pytorch.org/whl/cpu safetensors huggingface_hub tqdm numpy tensorboard gradio git+https://github.com/AbstractEyes/alephllm README.md markdown -– title: Beatrix 376M Splat Memory Lab emoji: colorFrom: blue colorTo: purple sdk: gradio sdk_version: 4.44.0 app_file: app.py pinned: false -– # Beatrix 376M — Splat Memory Lab Verified replication — only measured values. **Model:** AbstractPhil/alephllm-mini-beatrix-training mini-beatrix-3 (376M, 32 blocks) **Probe:** clean 14.75 bits / interfered 17.59 bits / delta +2.84 bits (measured) **Interpretation from code reading and article:** clean ~3 bits/digit near uniform 3.32, interfered retrieves planted value. **Mechanism:** - Write: Sp += k⊗v — running sum, never erased - Read: query·memory over two halves normalized by agreement mass - Addressing: sinh/cosh form per README - No softmax over positions **K-sweep (signal ID width):** K=8:0.49, 32:0.41, 128:0.43, 512:0.38 — weakly less interference, noisy due to concentration. **Tabs:** 1. Interference Probe 2. Encoder — real hidden states at any layer 0-31 3. Mechanism Demo — dist to mean ~0 proves additive Weights auto-download via load_trunk() on first launch (~720MB), CPU ~4GB RAM. Deploy: 1. Hugging Face – The AI community building the future. → Gradio SDK 2. git clone your Space 3. copy app.py, requirements.txt, README.md 4. git add . && git commit -m “initial” && git push First launch few minutes for weights, then cached. Same code you approved — no Madera claim, no Phil conversation, read described correctly.
discuss.huggingface.co
October 11, 2026 at 3:47 AM
Scaling Beatrix V3: a 376M byte-level model and the training
Thanks for running it, and for posting the generated write-up as is. A few things in it are right, one measurement is misread, and the verdict at the end answers a question nobody asked. In order: **What the write-up gets right.** The install and the loader are correct: `pip install git+https://github.com/AbstractEyes/alephllm` for the base package and `load_trunk(ckpt="weights/step_00245674.safetensors")` from `geolip.alephllm.arm_mount` are the real entry points, and the model takes byte ids 0 to 255. The quoted write line is also real: every block’s memory is a running sum of key-times-value outer products, and nothing in the sum is ever removed. We say the same in the article, with the same consequence drawn: the memory never erases, and that is why the repair work exists. One detail is off in the read line. The two halves are the positive and negative parts of a hyperbolic address, and the read divides a sinh-weighted sum by a cosh-weighted one, so the quoted numerator is only half of the rule and the sign of the second term lives in the keys. There is no softmax over positions, which is the point of the design, and cancellation between writes is indeed the only subtraction. **What the probe measured.** The “clean” prompt plants no code at all, so 14.75 bits for five digits is the model guessing digits at about three bits each, close to the 3.32 bits of a uniform guess over ten. The “interfered” prompt plants a different code, 99999, under the same key and then scores 12345. A model that retrieves the planted value will of course pay more bits on a contradicting one. The +2.84 bits is not noise leaking into the logits; it is the model partly fetching what was planted and being scored against something else. A standard softmax model pays that penalty harder, not softer, because it fetches the planted value more completely. Also worth keeping apart: the first pair of numbers in the write-up, 2.14 and 6.68, was written before anything ran; the second pair, 14.75 and 17.59, is the measured one. Only the second should be kept. The informative versions of the test are three, and we have run all three on the same checkpoint with paired trials inside real documents. They are public with their per-trial rows and the tool that produced them, so anyone can rerun them: test | what it asks | v3 release | a softmax control of the same recipe ---|---|---|--- plant the same code, ask at its cue d bytes later | does the first digit come back | 0.06 at the plant, 0.00 at 1,024 and 2,048 bytes | about 1.0 at the plant, 0.94 at 1,024, 0.95 at 2,048 four records with different keys, ask for one | is the right record chosen | chance | chance the same key written twice, “is now” the second time | which value comes back | neither | the last 0.43, the first 0.38 Row one is the real failure of the release model: it does not make the jump from a cue to a value stored behind text. Row two is not a splat-versus-softmax difference at all; neither model keys among competitors at this size and training. Row three is where an update test belongs, and the release model fails it too, in a different way from the control. **The toy proof.** The toy is correct about the toy: two writes with one key read back as their mean. It is also the stated design, not a discovery against it. What it does not model is the actual read rule above, the chunked scan it runs over, or the trained keys, so “proves a flaw” overstates it. The question the article asks is whether that additive memory needs an erase, a decay, or a lookup path beside it. That is what is being tested now. **The verdict.** “Go-kart versus semi” compares a 376M research model with a 4,096-byte context against an assistant product, which it was never offered as. It has no instruction tuning, no tool loop and no safety layer because none of those were the experiment. The thread is about memory mechanisms, and the interesting disagreements are there. **Where the repair stands, measured.** The current run adds a detachable softmax attention arm after four of the thirty-two blocks of the released checkpoint, with every trunk weight frozen, and trains only the arm on web text with a small share of planted-record rows. It starts silent, so at step zero the model is the release exactly. Read every thousand steps with the same battery as the table above, one arm and one seed so far: arm steps | first digit of a planted code at the plant | at 1,024 bytes | at 2,048 bytes | held-out bits per byte ---|---|---|---|--- 0, the release | 0.06 | 0.00 | 0.00 | 0.9879 1,326 | 0.20 | 0.02 | 0.01 | 0.9833 2,326 | 0.95 | 0.50 | 0.19 | 0.9829 3,326 of 4,000 | 0.94 | 0.60 | 0.36 | 0.9833 The fetch arrives in a phase change between the second and third read and its range extends at every read since, with no cost on held-out text. The two things that have not moved are the choice among competing records, still at chance, and the order inside a long repeated string, which the softmax control handles and this arm so far does not. Those look like separate mechanisms from the fetch, and they are the next tests. The reads, the per-step checkpoints and the arm alone are under `mini-beatrix-3-la2-6-12-20/` on the training repo; the battery tool and the nine reference reads are under `reports/byte_battery_2026-10-10/`, and the arm’s code is in the library at tag v0.10.15 with a short doc. If you want to run the probe again, the planted-code test with the same code at a chosen distance is the one that separates retrieval from interference, and the tool takes the checkpoint name and step as arguments.
discuss.huggingface.co
October 11, 2026 at 1:47 AM
Air to Ground Message:

Howdy. When select all on ckpt intephone hand set we get rapid busy signal and cannot make PA until after press reset. Then PA disturbs crew rest. Fas verified all handsets secure

Area: Grand Rapids, MI, USA
Type: Airbus A330-300
A: #a0555f2fe3b
F: #f622b83e079
October 7, 2026 at 10:11 PM
Air to Ground Message:

HOWDY. WHEN SELECT ALL ON CKPT INTEPHONE HAND SET WE GET RAPID BUSY SIGNAL AND CANNOT MAKE PA UNTIL AFTER PRESS RESET. THEN PA DISTURBS CREW REST. FAS VERIFIED ALL HANDSETS SECURE

Area: Grand Rapids, MI, USA
Type: Airbus A330-300
A: #a0555f2fe3b
F: #f622b83e079
October 7, 2026 at 10:11 PM
Hallo Steffen, hallo Markus, wieder eine klasse Folge!
Vielen Dank dafür 🤓👍
Markus hatte wohl eine Vorahnung zur Axt im Cockpit 🫣 schrecklich! Die verschließbare Ckpt-Tür hat aus meiner Sicht keinen Sinn und gehört wieder aus den Regeln entfernt. Das war ein FAA Schnellschuss nach 9/11. Ich teile
October 4, 2026 at 12:03 PM
Air to Ground Message:

Packs are putting out very weak airflow and barely cooling cabin and ckpt. Not a writeup but definetly worth looking at

Area: Washington DC, USA
Type: Bombardier CRJ900
A: #a20ef25ee7b
F: #f007a97a30d
October 1, 2026 at 9:48 PM
Air to Ground Message:

PACKS ARE PUTTING OUT VERY WEAK AIRFLOW AND BARELY COOLING CABIN AND CKPT. NOT A WRITEUP BUT DEFINETLY WORTH LOOKING AT

Area: Washington DC, USA
Type: Bombardier CRJ900
A: #a20ef25ee7b
F: #f007a97a30d
October 1, 2026 at 9:48 PM
ガタつき軽減で、さらに没入感アップ!レースゲームがもっと楽しくなる秘密兵器🏁✨
https://hb.afl.rakuten.co.jp/ichiba/55c89ea4.29b98f70.55c89ea5.eb2be847/?pc=https://item.rakuten.co.jp/shunte/ckpt-strutbar/
September 28, 2026 at 6:42 PM
uvr5.6のベータ版を入れてから、viperx1297はダウンロードセンターからインストール
reviveはファイルを
huggingface.co/pcunwa/BS-Ro...
から拾ってきて手動でインストール
bs_roformer_revive2.ckpt · pcunwa/BS-Roformer-Revive at main
We’re on a journey to advance and democratize artificial intelligence through open source and open science.
huggingface.co
September 22, 2026 at 4:59 PM